Content area
ABSTRACT
Plant phenomics has emerged as a critical bridge between genotype and phenotype, addressing a significant bottleneck in crop breeding and functional genomics studies. Hyperspectral imaging, a key technology in this field, has been instrumental in high‐throughput, non‐destructive phenotyping. Compared to other imaging technologies, hyperspectral imaging stands out for its continuous and fine spectral resolution, capturing subtle changes in plant biochemical and physiological states, which is essential for precise identification and analysis of plant characteristics. Recent advances in deep learning have further expedited hyperspectral data analysis, fostered multi‐omics research and enhanced our ability to integrate diverse datasets. Despite challenges in establishing standards of data acquisition and processing, a significant proposal has emerged for the scientific community to collaboratively build a vast hyperspectral database. Integrated with reducing the cost of hyperspectral sensors and promoting more open‐source analysis pipelines for hyperspectral data, these initiatives promise to lay the groundwork for robust big data analytics, potentially revolutionising plant research and breeding.
Introduction
The advent of plant phenomics has marked a significant shift in plant science, emerging as a crucial discipline for elucidating the intricate interactions between genotype, environment and phenotype [1]. It involves meticulous control, comprehensive collection, and systematic analysis of plant phenotyping data alongside relevant environmental factors. This discipline bridges the gap between genomics and observable plant traits, offering a holistic view of plant behaviour across diverse conditions [2].
The evolution of plant phenomics is driven by the need to overcome the limitations of traditional phenotypic assessment, which has become a barrier in the era of plant functional genomics and molecular crop breeding [1]. With the completion of whole genome sequencing of key crops, there has been a significant increase in demand for high-throughput, precise and non-destructive phenotyping technologies. In this context, hyperspectral imaging (HSI) has emerged as a pivotal technology, providing a non-invasive and detailed approach to capture the spectral signatures of plant tissues across a wide range of wavelengths [3]. HSI, which captures both spectral and spatial information across a wide range of wavebands simultaneously, has emerged as a nondestructive powerful tool in the field of precision agriculture, offering unique insights into various aspects of plant physiology and health based upon the interaction of light with plant tissues and their biochemical components [4]. Plant science research uses HSI not only for monitoring growth and analysing quality but also increasingly to investigate the genetic basis of plants’ environmental responses. Analysis of 547 articles in the Web of Science Core Collection from 2020 to 2024 (for details on the methodology, see Supporting Information S1: Note S1) reveals wheat, corn, and rice as the most studied staple crops in hyperspectral studies (Figure 1a,b), while soybean, oilseed rape, and tobacco lead among cash crops (Figure 1c). However, research on rice is significantly less prevalent than on wheat and corn. This disparity stems primarily from challenges in collecting HSI data within paddy fields and the limited availability of suitable agricultural robots. Furthermore, Unmanned Aerial Vehicles (UAVs)-based HSI data analysis in this environment adds significant complexity. Beyond agricultural systems, studies on forests (Figure 1d) and grasslands (Figure 1e) frequently lacked species-specific resolution, whereas grapes, apples and strawberries emerged as the predominant fruit crops investigated (Figure 1f).
[IMAGE OMITTED. SEE PDF]
In terms of applications, HSI was extensively utilised for assessing nitrogen status, chlorophyll content, biomass, water status, photosynthesis parameters, Leaf Area Index (LAI) and yield estimation. Research on abiotic and biotic stresses primarily focused on drought stress, invasive weeds, water stress, heavy metal stress and Xylella fastidiosa (Xf) infection (Figure 1g). While most HSI research has focused on plant growth monitoring, future studies should increasingly target stress resistance and its underlying genetic mechanisms. Furthermore, HSI found significant applications in detecting protein content in agricultural products and facilitating variety discrimination (Figure 1g). Despite these advances, critical gaps persist. Current research remains heavily skewed toward growth monitoring, underscoring the need for future studies to prioritise stress resistance mechanisms and their genetic regulators. Moreover, the absence of standardised protocols for HSI-based quality trait detection hinders reproducibility, limiting broader adoption and reliable cross-study comparisons in plant science.
This review meticulously charts the course of HSI's burgeoning role across various facets of plant research: from its principles and foundational applications in plants to the traditional and advanced methodologies of HSI data mining, to the cutting-edge innovations in HSI technologies. Furthermore, the rich landscape of data processing software and toolkits is systematically explored, highlighting their role in streamlining and enhancing the utility of HSI data. Lastly, the contributions and development of hyperspectral databases are examined, with these repositories now recognised as indispensable archives of spectral information. These interconnected developments collectively advance the operationalisation of HSI insights, bridging fundamental plant science with field-deployable solutions for optimised crop management and environmental monitoring.
Principles of Hyperspectral Imaging
HSI is an advanced technology for acquiring spectral information, capable of capturing the spectral data of objects across multiple bands, thus enabling precise material analysis and identification. Its working principle can be summarised as quantifying the interactions between light and plant tissues as well as biochemical components, capturing changes in the absorption, reflection, emission, or transmission of electromagnetic waves corresponding to the physiological state of the plants [5]. Typically, HSI applied in agriculture covers bands ranging from visible light to near infrared (NIR) to short-wave infrared (SWIR), capturing reflected or radiated light signals and integrating spectral (λ) and spatial (x, y) data into a three-dimensional data structure known as the hypercube, as shown in Figure 2a [6]. This technology adds a spectral dimension to the traditional two-dimensional images, forming a three-dimensional data cube, which can identify the composition and properties of substances.
[IMAGE OMITTED. SEE PDF]
From a hardware perspective, an HSI system consists of an imaging lens, spectral acquisition device, detector and signal processing components, which work together to ensure effective data acquisition and analysis. HSI typically employs various spectral dispersion methods, such as dispersion, interference and absorption, to acquire spectral dimension information. The choice of dispersion method determines the technical specifications and application fields of the HSI system [7].
HSI Systems Based on Dispersion Methods in the Imaging Spectrometer
Based on the methods for obtaining spectral information in the imaging spectrometer, HSI systems can be classified into four types: dispersive, narrow-band filtering, Fourier transform (interference principle) and computational reconstruction, as shown in Figure 2b [8].
Dispersive imaging spectrometers are currently the most mature form of all imaging spectrometers, and due to their high performance and environmental adaptability. As the name suggests, dispersive devices, including prisms, gratings, and their combinations, are core optical components that diffract incident composite light into different monochromatic lights, as shown in Figure 2b. The basic components of a dispersive imaging spectrometer include a slit, a collimator, a dispersive element, a focussing mirror and a detector [9].
Filter-based imaging spectrometers include narrow-band filter types, tunable narrow-band filter types and array filters. Narrow-band filters can effectively extract characteristic spectra from the radiated light of the target while providing high suppression of out-of-band stray light, forming the simplest type of imaging spectrometer, as shown in Figure 2b. Currently, mainstream tunable narrow-band filters include acousto-optic tunable filters (AOTF), liquid crystal birefringence tunable filters (LCTF) and Fabry–Perot filters [10]. Taking the acousto-optic tunable filter as an example, it utilises the principle of acousto-optic diffraction to dynamically select specific wavelengths and can adjust the wavelength at microsecond speeds. Array filter-based imaging spectrometers can be realised by coating thin films, meta-surfaces, and other filter arrays on the photosensitive surface of photodetectors [8].
Interferometric imaging spectrometers, generally referring to Fourier transform spectrometers, operate based on the wavelength interference effect with optical path differences [11]. The traditional Fourier transform spectrometer (i.e., Michelson interferometer) uses a moving mirror to produce different optical path differences. The beam with an optical path difference is recombined by the focussing mirror and received by the detector. By analysing the detected interferogram, the spectral and imaging information of the target can be obtained with extremely high spectral resolution, as shown in Figure 2b.
Computational reconstruction spectrometers introduce algorithmic reconstruction into the system. The incident spectrum is encoded through filter and then combined with an algorithm to decode and reconstruct, improving information utilisation and alleviating the trade-off between spatial resolution and spectral resolution, as shown in Figure 2b. In 2019, Yang et al. [12] designed a single-nanowire miniature spectrometer based on the response differences of nanowire materials under different conditions, creating an ultra-compact computational miniature spectrometer.
Imaging Spectrometers: Acquisition Methods for Hypercube
Based on the different ways of how to acquire the three-dimensional spectral data cube, imaging spectrometers can also be divided into point scanning, line scanning, spectral scanning, and snapshot types, as shown in Figure 2c. Scanning-type imaging spectral systems refer to systems that measure 1D or 2D slices of the data cube in a time sequence to build a 3D data cube. Snapshot-type imaging spectrometers [13], on the other hand, use dispersive elements or aperture diaphragms to enable the spectrometer to simultaneously obtain the three-dimensional data cube. Snapshot-type imaging spectrometers mainly include computational tomography imaging spectrometers and coded aperture computational imaging spectrometers.
Detectors and Their Operational Wavelength Ranges of HSI System
Another crucial component is the detector, which directly affects the operational wavelength range of the HSI system. The operating wavelength ranges of common detectors, including Complementary Metal-Oxide-Semiconductor (CMOS), InGaAs, HgCdTe, amorphous silicon (a-Si) and vanadium oxide (VOx), are shown in Figure 2e. Among them, the CMOS and InGaAs detectors are photovoltaic devices, with their photoresponse principle illustrated in Figure 2e. The HgCdTe detector exhibits both photovoltaic and photoconductive response mechanisms, with the photoconductive response principle shown in Figure 2e. Both a-Si and VOx detectors are photo-thermoelectric devices, with their photoresponse principle depicted in Figure 2e.
The response wavelength range of silicon detectors is limited to 0.2–1.1 μm, with a detectivity reaching up to 1013 cm·Hz1/2W−1. In recent years, significant efforts have focused on enhancing detector quantum efficiency and broadening their spectral response. n+/n-type silicon photodetectors based on black silicon can operate between 400 and 1600 nm [14]. Additionally, by introducing micro- and nanoscale light-trapping structures, effective light capture is achieved, resulting in an external quantum efficiency of 52% at 850 nm and 62% at 800 nm in silicon photodiodes [15]. Furthermore, a novel dual-band GeSn/Ge/Si detector, which vertically integrates PIN photodiodes, has been developed, enabling dual-band response in the near-infrared (0.75–1.4 μm) and short-wave infrared (1.4–2.5 μm) ranges, significantly broadening its operational spectrum [16]. Currently, various CMOS detector components with different resolutions are available from numerous domestic and international companies such as Sony, Hamamatsu, Gpixel Changchun Microelectronics and Brigates Microelectronics.
InGaAs is the most superior material for near-infrared detectors. Common InGaAs short-wave infrared detectors cover the wavelength range of 0.9–1.7 μm. By employing InP substrate removal techniques, the detector's cutoff frequency can be extended to 0.4 μm in the short-wave direction. Additionally, adjusting the In material composition allows for detection wavelengths to be extended up to 2.5 μm [17]. Quantum well-coupled In0.53Ga0.47As/In0.52Al0.48As quantum cascade detectors, designed based on band engineering, have demonstrated detection capabilities in the mid-wave infrared (5.4 μm) range [18]. Various InGaAs short-wave infrared detectors with resolutions such as 640 × 512 and 1280 × 1024 are currently offered by companies like Princeton Infrared in the United States., Sony in Japan, and the Shanghai Institute of Technical Physics in China.
The HgCdTe material allows for the adjustment of Cd composition to cover the entire infrared wavelength range. In the short-wave infrared (1–3 μm), mid-wave infrared (3–5 μm) and long-wave infrared (8–14 μm) atmospheric windows, its detectivity can exceed the level of 1011 cm·Hz1/2W−1 [19]. Currently, HgCdTe infrared detector components with resolutions of 640 × 512 and 1280 × 1024 pixels across different operational wavelength bands are available from various institutions, including Kunming Institute of Physics in China, North China Research Institute of Electro-Optics, Sofradir in France and Selex in the UK.
a-Si and VOx are the two mainstream materials for uncooled infrared detection. Typically, the operating wavelength range for amorphous silicon detectors and vanadium oxide detectors is 8–14 μm. Although the former is easier to produce on a large scale, the latter significantly outperforms it in terms of detection sensitivity and temperature measurement accuracy [20]. Vanadium oxide detectors are represented by companies like DRS, Raytheon, FLIR in the United States, BAE in the UK, SCD in Israel, and IRay Technology and North Guangwei Technology in China, while amorphous silicon detectors are represented by ULIS in France and Dali Technology in China.
Calibration of HSI System
HSI sensor calibration includes wavelength and radiometric calibration. Radiometric calibration of a hyperspectral camera system (sensor) ensures that the sensor’s output (digital numbers or raw data) is converted to absolute physical units (e.g., radiance in W/m2·sr·nm) with high accuracy. This process corrects for sensor-specific variations, environmental influences, and system noise. The wavelength calibration ensures that a spectral device (e.g., spectrometer, hyperspectral camera) accurately maps measured signals to their corresponding wavelengths. The calibrated spectral imaging instruments can output accurate wavelengths and radiances.
Extracting, analysing, and utilising information from hyperspectral images (HSimg) remains another significant challenge. Traditional imaging spectrometers use displacement platforms and area array detectors to capture spectral data along the spatial dimension of the target in one line per frame. By extracting spectral dimension data from all frames of the area array detector and stitching them together, a reconstructed three-dimensional spectral data cube is obtained. The data reconstruction method is illustrated in Figure 2d.
The Future of Hyperspectral Imaging: Miniaturised Systems
HSI is rapidly evolving from traditional laboratory-based systems toward compact, field-deployable solutions, driven by advancements in miniaturised spectrometers, UAV integration, and chip-scale optical technologies. Modern HSI systems now leverage deep learning (DL) for real-time spectral analysis, enabling applications ranging from precision agriculture and environmental monitoring to medical diagnostics and industrial quality control. The integration of super-resolution techniques and adaptive optics is pushing beyond conventional resolution limits, while the emergence of hyperspectral video (4D imaging) opens new possibilities for dynamic spectral-spatiotemporal analysis. These developments, combined with edge computing capabilities, are transforming HSI from a specialised analytical tool into a versatile, real-time sensing solution with broad industrial and scientific impact.
Applications of Hyperspectral Imaging in Plant Science
Plant Growth Monitoring
Precise assessment and management of plant growth are paramount for the optimisation of fertiliser usage, amplification of plant yield potential, and mitigation of adverse environmental ramifications associated with over-fertilisation. This includes monitoring nutrient status, assessing growth stages, forecasting yield, and identifying biotic and abiotic stresses.
Agricultural producers often pay attention to the nutrient status of their plants. Leaf chlorophyll content (LCC) serves as a crucial indicator of nutrient status. It has been reported that the most common method for LCC assessment is based on hyperspectral vegetation indices (HVIs) [21], including the Normalised Difference Vegetation Index (NDVI), the Triangular Greenness Index (TGI), and the MERIS Terrestrial Chlorophyll Index (MTCI), etc. Another alternative method for determining LCC involves estimating it based on its strong correlation with the SPAD value that has also been effectively evaluated via HSI data [22]. Furthermore, HVIs are frequently employed to evaluate the levels of nitrogen and water content, which are crucial factors in determining the photosynthetic capacity of plants [23]. In recent years, ML algorithms have demonstrated their superiority in capturing the non-linear relationships between spectral data and plant nutrient status, encompassing nitrogen, phosphorus, potassium and other essential nutrients [24]. For instance, Furlanetto et al. developed a non-destructive hyperspectral framework using ten machine learning (ML) algorithms (e.g., partial least squares regression [PLSR], support vector machines [SVM], and genetic algorithm [GA]-optimised methods) to predict potassium (K+) in soybean leaves, achieving high accuracy (R2 ≈ 0.88) through wavelength selection in critical spectral regions (blue, green, red-edge, NIR), enabling early K+ monitoring [25].
Plant growth is predominantly evident in their above-ground parts, where above-ground biomass (AGB) and LAI serve as crucial metrics for assessing plant growth stages [26]. At the satellite scale, Tagliabue et al. successfully implemented a hybrid retrieval scheme using PRISMA hyperspectral satellite imagery to estimate key crop traits, with the LAI achieving a high accuracy R2 value of 0.84 [27]. At the UAV scale, Xu et al. successfully utilised models combining HVIs with corresponding-band texture (VI-CBT) and full-band texture (VI-FBT) derived from UAV HSimg to dynamically estimate rice AGB with improved monitoring accuracy at different growth stages [28]. At the proximal hyperspectral scale, Feng et al. addressed the critical challenge of spectral saturation in dense canopies by developing a stratified VGC-AGB model. This framework synergises proximal hyperspectral data with DL-inverted biophysical parameters (LAI and dry mass content), achieving multi-stage potato AGB estimation with field-level precision with R2 value of 0.853 and RMSE value of 751.12 kg/ha, thereby establishing a cross-scale sensing paradigm from plot to regional monitoring [29]. Additionally, AGB and LAI serve as critical biophysical parameters for assessing crop productivity, as they directly reflect photosynthetic capacity and resource allocation efficiency. Their strong correlation with final yield makes them indispensable indicators of yield forecasting models, particularly when integrated with hyperspectral remote sensing technologies [30]. For example, Kayad et al. utilised the PROSAIL model to retrieve LAI from aerial hyperspectral imagery, which served as a basis for estimating corn grain yield (R2 = 0.69) [31].
In modern agriculture, biotics not only threaten the vitality and resilience of crops but also have profound implications for food security, sustainability, and agricultural productivity globally. As reported in the research conducted by Jin et al. [32], from 2004 to 2023, the top five crops that have been the focal point in the utilisation of HSI for disease monitoring include wheat, rice, olives, potatoes and oil palms. Additionally, the diseases that are most recurrently highlighted in these papers are blight, rust, mildew, rot and spot. These articles encompass the use of HSI for early detection, identification, characterisation, and classification of pests and diseases. Notably, Jin et al. demonstrated that integrating supervised classification methods (e.g., SVM, RF, decision trees and threshold-based techniques) with physically grounded biochemical and structural approaches optimises spectral data utilisation, significantly advancing HSI as a robust tool for dynamic monitoring plant diseases [32]. For instance, Feng et al. proposed an ML-based method for rice blast identification, achieving high classification accuracy with overall accuracy (OA) and Kappa coefficients of 97.21% and 96.55% for the RF model, respectively [33]. Similarly, the PLS-DA model has proven effective in disease classification tasks. Xuan et al. utilised a PLS-DA model with a fused dataset of effective wavelengths and texture features extracted from HSI, achieving a classification accuracy of 91.4% for the early diagnosis of wheat powdery mildew caused by Blumeria graminis [34]. In recent years, DL methods have shown significant advantages in HSI analysis and plant disease monitoring. For instance, Wang et al. [35] developed a Spectrum Transformer Network (STNet) that leverages self-attention mechanism (AM) to model global spectral dependencies, addressing challenges in high-dimensional HSI data and limited labelled samples. By pretraining on unlabelled pixel-level spectra and fine-tuning with labelled data, STNet achieved superior performance in tomato bacterial wilt severity identification (91.93% accuracy, 0.9309 F1-score), outperforming SVM-, RF-, and CNN-based models [36]. This highlights DL's ability to automate feature extraction and enhance early disease detection, offering scalable solutions for precision agriculture.
However, abiotic stresses, including nutrient deficiency, drought, waterlogging, salinity, heavy metals, herbicides, extreme temperatures, may be more damaging than diseases and pests infectious on plant yields, which have been extensively detected via HSI [37]. As comprehensively reviewed by Berger et al. [37] and Paulus et al. [38], ML algorithms, such as SVM, RF, PLSR and Genetic Algorithms (GA), are extensively applied to classification and regression tasks in stress diagnostics. For example, SVM and RF classifiers achieve up to 97% accuracy in categorising stress types (e.g., heavy metal contamination, nutrient deficiency) and severity levels by decoding spectral reflectance patterns under controlled conditions. PLSR models demonstrate exceptional performance in quantifying physiological parameters (e.g., chlorophyll content, water status) under drought or salinity stress, with validation R2 values ranging from 0.70 to 0.93. Feature selection techniques, including Successive Projections Algorithm (SPA) and Competitive Adaptive Reweighted Sampling (CARS), enhance model interpretability by isolating stress-sensitive spectral bands, such as 530–630 nm for chlorophyll degradation. And DL approaches, particularly CNN, are leveraged for high-dimensional HSimg analysis, enabling pixel-level stress mapping and early detection of subtle biochemical changes. For example, Chu et al. employed Vis/NIR HSI with shallow CNN to evaluate herbicide stress in wheat, achieving 96% accuracy in herbicide type and 80% in stress level classification after 48 h, showing the potential of HSI for rapid detection of herbicide effects [39]. Notably, most abiotic stresses are inextricably linked to farmland contamination. Farmland pollution, often manifested as heavy metal contamination and pesticide residues, accumulates in the soil and is absorbed by vegetation, thereby impeding soil fertility and plant growth. HSI is capable of not only tracking heavy metal stress in plants but also of assessing the soil conditions. For instance, Yao et al. proposed a stacking model that integrates GaoFen-5 hyperspectral data and geographical environmental factors to predict Soil Heavy Metal Concentrations (SHMC), achieving a significant improvement in predictive accuracy, particularly for Cd and As with a 52% and 48% decrease in RMSE, respectively [40].
In summary, HSI serves as a transformative tool for precision agriculture, enabling multi-tiered plant growth management: (1) optimising nutrient use efficiency to reduce over-fertilisation, (2) detecting biotic/abiotic stresses for timely interventions, (3) tracking plant phaenological stages to optimise harvest timing and yield potential and (4) monitoring soil health to mitigate contamination risks. By translating spectral data into actionable insights, HSI enhances plant resilience while minimising environmental footprints. As advancements in sensor portability and AI-driven analytics progress, this technology will increasingly empower stakeholders to harmonise agricultural productivity with ecological stewardship, addressing global challenges such as climate variability and food security.
Plant Breeding and Genetic Improvement
Plant breeding is a critical component in the global effort to ensure food security and adapt agriculture to the challenges posed by climate change. As the world’s population continues to grow and the effects of climate change become increasingly evident, the importance of developing plant varieties that can withstand environmental stresses and maintain productivity cannot be overstated. The integration of modern breeding techniques, such as genomic selection and high-throughput phenotyping, with artificial intelligence and ML algorithms has the potential to significantly accelerate the process of plant improvement [41]. HSI has been instrumental in realising the key breeding objectives of achieving high yields, enhancing quality, and improving resistance. Moreover, HSI has been crucial in deciphering the complex interactions between genetic factors and environmental conditions, providing valuable insights that guide the development of plant varieties that are better equipped to face the dynamic demands of modern agriculture.
Building on these objectives, HSI has demonstrated significant potential in identifying superior varieties that meet the criteria for high yield, quality, and resistance. Joynson et al. employed HSI and GWAS to identify candidate genes associated with photosynthetic capacity that affects yield in wheat, notably GUN5, which plays a crucial role in chlorophyll synthesis by encoding magnesium chelatase, and SWEET4/5, a bidirectional sugar transporter implicated in regulating LCC [42]. Furthermore, it was identified that the OsmtSSB1L gene was associated with rice grain protein content by integrating HSI data, deep convolutional Generative Adversarial Networks (DCGANs) and GWAS [43]. These studies underscore the significant potential of HSI-assisted breeding in developing plant varieties with improved yield and quality. Moreover, HSI has proven advantageous in the selection of resistance genes. Zhang et al. explored the genetic underpinnings of salt stress response in rapeseed by utilising GWAS and linkage analysis on traits derived from RGB and HSI data, leading to the discovery of two unreported genes, BnCKX5 and BnERF3, that were experimentally verified to play a role in regulating the salt stress tolerance of rapeseed [44]. 15 i-traits, collected by RGB imaging, HSI, and X-ray computed tomography (CT), were identified as potential markers for corn drought tolerance breeding [45]. Furthermore, by combining ML, HSI, GWAS, and RNA-seq, Yang et al. have efficiently and accurately identified two candidate genes—COS02g_02406 for Cd resistance through isoflavonoid biosynthesis and antioxidant defence, and COS06g_03984 for Cd absorption via promoting Cd2+ binding to ETR/ERS receptors—in jute [46].
Notably, the interplay between genotype and environment, known as GxE interactions, is a critical consideration in the field of plant breeding. HSI, a vital tool for capturing phenotyping data, has also emerged as a significant contributor in this area of research. For example, HSI coupled with environmentally sensitive fluorophores (ACDN/LAURDAN) and phasor analysis was employed to quantify salinity-modulated changes in biomolecular condensate hydration states and membrane lipid packing, elucidating environment-dependent molecular remodelling [47]. What's more, Corbin et al. utilised hyperspectral reflectance to uncover genetic variations and environmental interactions in Populus fremontii, achieving high accuracy in identifying population identity and revealing the heritability of spectral indices, GxE interactions, and directional physiological adjustments, thereby highlighting the effectiveness of hyperspectral data in monitoring genetic and phenotypic responses to climate change [48].
In summary, the deployment of HSI in plant breeding has significantly bolstered our capacity to cultivate plants that exhibit enhanced resilience, higher productivity, and greater adaptability to environmental pressures. As this technology continues to evolve, its integration into plant breeding programs is set to become a cornerstone in the quest for food security and the advancement of sustainable agricultural practices. This approach is particularly crucial in the context of the ever-changing challenges posed by climate variability and the urgent need to feed a growing global population. The future of agriculture hinges on such innovative technologies, which offer a beacon of hope for enhancing crop performance and ensuring the long-term viability of our food systems.
Plant Quality Evaluation and Grading
Quality assessment and grading of plants are of paramount importance in agriculture and food industries. They ensure that the produce meets the required nutritional and safety standards, facilitate trade by setting industry benchmarks, and ultimately contribute to consumer satisfaction and health. With the growing demand for high-quality food and the need for precision agriculture, the role of non-destructive and rapid assessment techniques has become increasingly significant. HSI stands out as a powerful tool in this context, offering the potential for real-time, in-depth analysis of plant quality attributes.
As living standards rise and the focus on dietary health intensifies, there is a growing public interest in the quality and safety of food. This concern spans the entire spectrum of agricultural product management, from harvesting and processing to storage and distribution. HSI, a powerful and non-destructive technique, is widely applied in assessing the quality of food and agricultural products including rice, wheat, corn, soybeans and various fruits and vegetables, measuring both basic quality parameters such as protein, moisture, fatty acids and micronutrients, as well as complex traits like maturity, ripeness, defect detection, physiological disorders, mechanical damage and sensory quality [49]. ML models, especially PLSR and SVM, are more widely used in HSI-based quality evaluation of agricultural products [50]. For instance, Belgiu et al. developed a novel approach utilising PRISMA and Sentinel-2 satellite imagery combined with Two-Band Vegetation Indices (TBVIs) and PLSR models to predict nutrient concentrations in crop grains, achieving notable accuracy with R2 values up to 0.73 for phosphorus in soybean grains and 0.67 for potassium in wheat grains using Sentinel-2 images [51]. While HSI has traditionally been used alongside chemometrics, recent breakthroughs in DL have transformed the analysis of spectral data for food and agricultural products, with the combination of HSI and AI substantially enhancing the evaluation of grain quality [52]. Wang et al. proposed an ACLSTM model that integrated AM, CNN, and long short-term memory (LSTM) with HSI for the quality evaluation of Coix seed (CS), achieving robust prediction performance with RPD values all higher than 3.0 for the quality indicators of oil, protein and starch [53]. Additionally, a CNN model, integrated with HSI, was employed for the rapid and non-destructive quality assessment of Canarium indicum nuts, achieving an overall accuracy of 93.48% in peroxide value (PV) estimation, which is crucial for evaluating the degree of oil oxidation and predicting the shelf life and quality of nuts [54].
As consumers prioritise the nutritional value of grains and farmers focus on aspects like seed vigour and germination rates, it becomes crucial to assess and categorise seed quality during seed distribution to meet the diverse needs of both parties effectively. To address the challenge of sample imbalance in seed viability assessments, Wu et al. developed a non-destructive testing technique for rice seed viability using a Deep Convolutional Neural Network (DCNN) with weighted loss and HSI, achieving exceptional accuracy and Macro F1 scores of 97.69% and 97.42%, respectively [55]. Similarly, Qi et al. demonstrated that Spectral Angle Mapper GAN, when combined with real spectral data, achieved near-perfect classification accuracy (99%–100%) for rice seed viability detection across multiple varieties, while effectively addressing the challenge of limited naturally-aged samples through optimised spectral data augmentation in the 900–1700 nm range [56]. Malik et al. developed a CNN model using HSI data to predict the quality of gypsum tofu from soybean seeds, achieving a high accuracy rate of 96%–99% for categorising new soybean seeds into distinct quality classes [57]. Additionally, A 3D hyperspectral full-surface imaging fruit grading system was developed, utilising a virtual volume intersection algorithm and texture technology, which achieved high accuracy in predicting pear volume (R2 = 96.18%) and mass (R2 = 98.18%), and effectively classified pears of different qualities with an accuracy of 95.33% utilising a ResNet-18 network [58]. Furthermore, a convolutional autoencoder-SVM model was employed for quality grading of Pleurotus eryngii during post-harvest storage, demonstrating superior performance with an accuracy of 91.58%, an F1 score of 91.36%, a precision of 89.65%, and a recall of 90.60% [59].
In summary, these advancements underscore the pivotal role of HSI in revolutionising agricultural quality assessment and grading. For regression-driven tasks such as nutrient quantification (e.g., protein, moisture), ML methods such as PLSR and SVM remain highly effective due to their interpretability and robustness with spectral data. Conversely, classification-driven analyses—requiring high-dimensional feature extraction and pattern recognition—benefit significantly from DL architectures like CNN, which excel at automating hierarchical feature learning from complex hyperspectral datasets. Nevertheless, classical ML classifiers, including SVM and PLS-DA, still hold value in situations with restricted training data or limited computational resources, demonstrating the complementary strengths of traditional and modern approaches in agricultural analytics.
Challenges and Synergistic Solutions
HSI has emerged as a widely used tool in plant science, enabling high-throughput, multi-scale phenotyping and advancing our understanding of above-ground plant physiology. However, several limitations remain: (1) Challenges in situ application, particularly in hydromorphic environments such as rice paddies, where water-mediated spectral interference affects data fidelity; (2) Root system phenotyping gaps, with subterranean structures remaining largely uncharacterised due to sensor accessibility constraints; and (3) The absence of standardised protocols for different studies. To overcome these challenges, future research could explore: (1) integration of HSI with LiDAR-derived 3D structural models and multi-omics data for enhanced trait analysis; (2) advanced computational methods, including explainable artificial intelligence (XAI) for physiologically interpretable feature extraction and edge computing for real-time field processing; and (3) developing scalable infrastructure through cost-optimised sensors, open-source algorithms, and community-driven standardisation to broaden HIS application.
Methods of Hyperspectral Data Mining
Despite its widespread success across diverse applications, the extraction, analysis, and effective utilisation of information from HSimg continue to pose significant challenges. A robust, multi-step data processing strategy is imperative for this task (shown in Figure 3 and Table 1). Within this strategy, data preprocessing emerges as a critical preliminary step, with a substantial effect on the fidelity of subsequent analyses. This phase encompasses two fundamental components: noise reduction and dimensionality reduction.
[IMAGE OMITTED. SEE PDF]
TABLE 1 Representative methods of HSI data mining.
| Processing target | Method type | Method subtype | Representative methods | Reference |
| Denoising | Filter-based | — | MNEM | [60] |
| Morphological + NSST + bitonic | [61] | |||
| MNF-BM4D | [62] | |||
| Optimised bilateral filter | [63] | |||
| Model-based | SDTVLA | [64] | ||
| Inter-block and intra-block sparse estimation framework | [65] | |||
| Non-negative tucker decomposition | [66] | |||
| NLR-CPTD | [67] | |||
| Learning-based | Supervised | Plug-and-play DLD | [68] | |
| SSGN | [69] | |||
| HIS-DeNet + residual learning strategy + extended convolution + multi-channel filtering with CNN | [70] | |||
| 3D-DCNN | [71] | |||
| Alternating direction three-dimensional quasi-recurrent neural network | [35] | |||
| Unsupervised | DIP algorithm using 3DCNN | [72] | ||
| S2DIP | [73] | |||
| DS2DP | [74] | |||
| Reduction dimension | Data compression | Traditional | TFEMPR | [75] |
| One-pass lossless compression framework | [76] | |||
| Based on two spatial, integer-to-integer biorthogonal graph filter banks lossless compression | [77] | |||
| DL | M2H-Net | [78] | ||
| A compression method based on neural network and clustering strategy | [79] | |||
| Feature extraction | Traditional | SMLAD | [80] | |
| LADA/2DLADA | [81] | |||
| SALDA | [82] | |||
| ASVMK | [83] | |||
| RF-OAK | [84] | |||
| KS-LEE | [85] | |||
| SNML | [86] | |||
| RDLVM | [87] | |||
| LPPAE | [88] | |||
| DL | Network based on DL and multi-level feature extraction | [89] | ||
| Network based on CNN-based encoder-decoder feature extraction | [90] | |||
| S2ADet | [91] | |||
| cs2GAN-FE | [92] | |||
| 3D-CSSEAN | [93] | |||
| New finite element method based on DL | [93] | |||
| Modelling | Classify | ML | Tensor algebra operation + ML algorithm | [94] |
| Nonlinear classification model whose weights satisfy the rank-1 regular decomposition property | [95] | |||
| Unsupervised classification method based on robust popular matrix factorisation | [96] | |||
| DL | DRCNN | [96] | ||
| SWFormer | [97] | |||
| Smote algorithm + CNN | [97] | |||
| CDA | [98] | |||
| SR-assisted HSimg classification deep network architecture based on classification loss SRCL | [99] | |||
| Regression | ML | FSR | [100] | |
| RF and LightGBM | [101] | |||
| DL | CNNs + deep autoencoders | [101] | ||
| DeepRWC | [102] | |||
| Temporal-spatial-spectral information fusion and extraction + CNN | [103] |
Spectral Pre-Processing
Spectral pre-processing is a critical step in HSI data analysis, where various techniques are employed to enhance the quality of spectral data and extract meaningful information. To safeguard the integrity and quality of spectral data, it is imperative to identify and eliminate spectral outliers that may arise from instrumental errors, environmental factors, or sample inconsistencies.
First-order and second-order derivatives are used to emphasise small variations in spectral data by removing baseline shifts and linear trends, respectively. They help in identifying narrow absorption features that might be obscured by a sloping baseline [104]. Savitzky–Golay (SG) smoothing is a technique that convolves the data with a Savitzky–Golay filter, which fits a polynomial to a sliding window of the data points, thereby reducing noise while preserving the shape and width of the signal [105]. Multiplicative Scatter Correction (MSC) is a scatter-corrective method designed to reduce variability between samples due to scatter and baseline shifts. It is particularly useful when dealing with spectral data that exhibit multiplicative scatter effects. Standard Normal Variate (SNV) is another scatter-corrective technique that normalises the spectral data by subtracting the mean and dividing by the standard deviation of each spectrum, making it less sensitive to scatter effects [106].
Fourier Transform (FT) is a mathematical technique that transforms a signal from its original domain (often time or space) to a representation in the frequency domain. It is used in spectral analysis to identify periodic components within the data. Wavelet Transform (WT), on the other hand, is a more advanced form of Fourier analysis that allows for the analysis of signal components at different scales or resolutions, making it particularly useful for non-stationary signals and image processing [107]. These techniques, along with the others mentioned, are chosen based on the specific characteristics of the spectral data and the goals of the analysis. By employing these methods, researchers can effectively enhance the quality of their spectral data, making it more accurate and meaningful for classification, regression, and other analytical tasks.
Denoising
Due to the limitation of physical and optical mechanism factors, the imaging chain of HSimg will be affected by many factors such as atmospheric environment, illumination and sensor [108]. Moreover, when the photon reaches the camera, the noise inside the instrument also degrades the quality of the image. Denoising techniques are crucial for mitigating the effects of atmospheric radiation, thermal noise, quantisation noise, and scattering noise, thereby enhancing the signal-to-noise ratio and analysis efficiency of HSI data [108]. Current HSI denoising methods are categorised into three main approaches: filter-based, model-based and DL-based methods.
Filter-Based Denoising Method
The distinct frequency distributions of authentic image signals and noise underpin the principle of filter-based denoising in HSI. Different filters can remove noise of different frequencies. Notably, the SG smoothing and median filtering techniques are adept at eliminating noise within the spectral domain for each pixel, demonstrating robust denoising and image reconstruction capabilities. However, the complexity and diversity of noise in HSI, spanning both spatial and spectral domains, often surpass the capabilities of conventional filters. To address this, several optimised filtering methods have been introduced. For instance, Xue et al. developed a Mixed Noise Estimation Model (MNEM) by integrating a Gaussian prior denoising model with filtering and the Sobel operator, enhancing the effectiveness of HSI noise estimation while preserving details and edge features [60]. Goyal et al. proposed a three-stage denoising process that leverages morphological operations, Non-Subsampled Shearlet Transform (NSST), and the bitonic algorithm, capitalising on the strengths of each to reduce noise and prevent the blurring of edges and gradients [61]. The Matching and 4D filtering (BM4D) algorithm, a benchmark in HSI denoising, was enhanced by Xu et al. through the Minimum Noise Fraction (MNF)-BM4D method, which applies guided filtering after MNF rotation to more effectively suppress both spatial and spectral noise [62]. Asokan et al. also contributed an optimised bilateral filter that employs natural heuristic optimisation algorithms to dynamically adjust parameters, ensuring optimal image denoising and edge detail preservation [63].
Model-Based Denoising Method
Model-based denoising in HSI fundamentally aims to eliminate mixed noise by leveraging the intrinsic characteristics of HSI. Conventional one-dimensional or two-dimensional denoising techniques, such as block matching and three-dimensional filtering (BM3D) and weighted nuclear norm minimisation (WNNM) [109], frequently overlook the inter-spectral correlations, potentially diminishing image quality. To address this, Sun et al. proposed a new spectral difference-induced total variation and low-rank approximation (SDTVLA) method, leveraging Spectral Difference Transform to project HSI data into the Spectral Difference Space, effectively restructuring and reducing mixed noise, especially sparse patterns like stripes or dead lines across bands [64]. In order to pursue hyperspectral and spatial fidelity, Gao et al. developed a framework that leverages inter-block and intra-block sparse estimation, exploiting the compositional and spatial-spectral structures within spectral images, and employing spatial alignment of spectral Laplacians (SASL) and local spectral low-rank (LSLR) properties to achieve spatial domain denoising without spectral accuracy loss [65]. The research shows that the tensor-based denoising method can preserve the intrinsic correlation of the HSimg structure and has a good denoising effect [110]. Tensor-based denoising methods, as demonstrated in research, are effective in preserving the intrinsic correlations of HSI structures. Bai et al. introduced a non-negative tucker decomposition method that utilises non-local similarity, significantly enhancing the quality of degraded HSI by processing three-dimensional patches as third-order tensors [66]. Additionally, Xue et al. proposed a Non-Local Low-Rank Regularised CANDECOMP/PARAFAC (CP) Tensor Decomposition (NLR-CPTD) model, which harnesses the Global Correlation across the Spectrum (GCS) and Nonlocal Self-Similarity (NSS) features of HSI to effectively denoise tensors composed of non-locally similar blocks [67].
Deep Learning-Based Denoising Method
The advancement of DL has revolutionised the approach to HSimg denoising, eliminating the need for costly optimisation and reliance on artificial priors. Recent work demonstrates this synergy through a plug-and-play Deep Low-rank Decomposition (DLD) model embedded within an optimisation framework, which effectively incorporates traditional priors while overcoming the limitations of hand-crafted constraints. This hybrid approach not only provides theoretical convergence guarantees but also achieves superior performance in tasks such as hyperspectral denoising and spectral compressive imaging, outperforming state-of-the-art methods by balancing generalisation and adaptability to complex noise distributions [68]. HSimg also contends with substantial mixed noise within the spectral domain. To address this, a spatial-spectral gradient network (SSGN), a hybrid noise reduction network, has been proposed. It is predicated on fully cascaded multi-scale DCNN and utilises spatial and adjacent spectral data concurrently. The network harnesses structural features across spatial dimensions to isolate noise and integrates spectral ancillary information for effective noise reduction, thereby mitigating spectral distortion across the entire framework [69]. The HSI-DeNet architecture, proposed by Chang et al., integrates residual learning strategies, extended convolution, and multi-channel filtering with CNNs, and further incorporates an adversarial framework [70]. This approach is robust and effective for managing mixed noise in hyperspectral imagery and enhancing the restoration process. Additionally, considering the spatial correlation and spectral correlation of HSI, Dong et al. have introduced a 3D-DCNN network, predicated on a separable filtering strategy. This network decomposes three-dimensional filtering into two-dimensional spatial and one-dimensional spectral components, addressing the spectral correlation of hyperspectral imagery while significantly reducing the network’s parameter count and enhancing denoising performance [71]. Wei et al. proposed an alternating direction three-dimensional quasi-recurrent neural network. This network substitutes two-dimensional convolution blocks with three-dimensional quasi-cyclic units and introduces an alternating direction structure to effectively model spatial-spectral correlations [35].
Supervised learning methods typically necessitate extensive datasets for training. However, HSimg acquisition is costly, and the availability of labelled samples is limited. Thus, employing a minimal number of samples for model training presents a significant challenge. Oleksii et al. have extended the unsupervised Deep Image Prior (DIP) framework, originally developed by Lempitsky et al., to HSimg processing. Their approach adapts the DIP algorithm using 3DCNN for HSimg denoising, inpainting, and super-resolution [72]. Additionally, Luo et al. optimised the DIP algorithm, introducing the S2DIP algorithm tailored for hyperspectral mixed noise removal. This algorithm leverages the depth prior of an unsupervised CNN to generate clean HSimg and combines the spatial-spectral total variation (SSTV) regularisation term with the minimisation norm to capture clean image spatial-spectral local smoothing and noise sparsity, thereby enhancing the DIP's mixed noise removal capabilities [73]. To tackle the substantial parameter learning challenges of the DIP algorithm in HSI data processing, Miao et al. proposed the DS2DP lightweight denoising framework [74]. This framework, based on the linear mixed model, ‘unwraps’ the spectral and spatial domains of hyperspectral imagery, learning the depth prior of the abundance map and endmember spectral features independently, reducing network parameters and complexity while improving performance.
While numerous high-performance denoising tools are available, achieving a high signal-to-noise ratio and high-quality images requires more than just tool optimisation. Analysing and processing the original image is equally crucial. A focus on the distribution of internal image information and noise, in conjunction with high-performance denoising tools can facilitate the acquisition of high-quality images and enhance the precision of data analysis outcomes.
Dimension Reduction
Although denoised data exhibit a superior signal-to-noise ratio and enhanced image quality compared to the original spectral data, they are not immediately suitable for direct analysis and research. This limitation stems from the high redundancy inherent in HSI technology, which captures a multitude of spectral bands during image acquisition, many of which are highly similar and thus contain redundant information. To address this issue, it is essential to apply suitable data compression and feature selection techniques to diminish the data's dimensionality and computational demands.
Data Compression
Hyperspectral data compression is essential for managing the large data volumes resulting from multiple spectral bands, which poses technical challenges in real-time applications, data analysis, and storage. Gundogar et al. introduced the TFEMPR method, a multidimensional array decomposition-based approach that employs a statistical recursion, eliminating the need for matrix inversion and relying solely on the input data's structure. This method is noted for its efficient and superior image reconstruction capabilities at low bit rates [75]. While lossy compression algorithms offer robust performance, they can introduce data distortion, which may hinder their use. Conversely, lossless compression techniques are advantageous for preserving the integrity of the original data. Carpentieri et al. proposed a one-pass lossless compression framework predicated on the predictive paradigm [76]. This framework employs a three-dimensional predictor tailored for hyperspectral data and integrates a data hiding strategy based on prediction error correction (MPE), enabling simultaneous lossless data labelling and compression. The receiver can then accurately reconstruct the original image. Tzamarias et al. presented two spatial, integer-to-integer, biorthogonal graph filter bank transformations, converting a lossy compression method into a lossless one. This approach reduces the computational complexity of image compression and achieves higher performance [77].
The advent of DL in computer vision has unlocked its potential for compressing high-dimensional data. The M2H-Net framework employs a modified U-Net architecture with residual connections and multi-scale convolutional blocks (1 × 1 and 3 × 3 kernels) to achieve accurate HSimg reconstruction from limited multispectral inputs (RMSE 0.010–0.016), demonstrating significant potential for HSimg compression by enabling 90%–97% data reduction while preserving spectral fidelity and classification accuracy (> 95%), as evidenced by its consistent performance across different sensor platforms and imaging conditions [78]. Mijares i Verdú et al. introduced a neural network and clustering-based compression method that clusters frequency bands to prevent the neural network from becoming overly complex due to an excessive number of input bands. The method also incorporates distance adaptive normalisation to mitigate the checkerboard effect in images with low variance and high depth. This network is compatible with the embedded implementation in airborne hardware, significantly reducing the computational costs and enhancing the efficiency of data transmission and analysis [79].
Feature Extraction
Feature extraction is a pivotal component of data dimensionality reduction, enabling the selection of salient feature information to support subsequent data processing. ML, a potent tool for spectral feature extraction, encompasses both traditional linear and nonlinear methods, differentiated by the nature of the mapping functions within the feature space.
Traditional Linear Methods
The classical linear feature extraction methods are designed to sample data exhibiting linear structures and extract image features through linear transformations. Principal component analysis (PCA) converts correlated variables into a set of orthogonal principal components, thereby decorrelating them. Canonical correlation analysis (CCA) identifies linear relationships between pairs of variables within a multidimensional dataset through multivariate statistical techniques [111]. Linear discriminant analysis (LDA) facilitates both classification and dimensionality reduction by projecting the most discriminative features onto an optimal linear subspace. These traditional approaches have not only achieved significant success in data dimensionality reduction but have also inspired algorithmic enhancements. For instance, the saliency-based multilabel LDA (SMLAD) [80], which is based on LDA, assigns weights to each data point through saliency estimation, addressing the issue of class imbalance in multi-label LDA scenarios. Locality adaptive discriminant analysis (LADA) and its two-dimensional variant (2DLADA) integrate local structure learning into the discriminant analysis framework, enabling the extraction of underlying data structures. This approach not only diminishes noise but also circumvents issues related to overfitting and the challenges posed by small sample sizes [81]. Self-weighted adaptive locality discriminant analysis (SALDA) integrates graph learning concepts into LDA, adaptively learning the optimal subspace from the original data to capture neighbourhood relationships and local information [82].
Nonlinear Methods
While the above methods have been successful in linear feature extraction, they may not be well-suited for capturing the nonlinear structures inherent in hyperspectral data. Consequently, there is a growing demand for the development of nonlinear feature extraction methods. Nonlinear dimensionality reduction algorithms are categorised into two primary approaches: the kernel function method and manifold learning, based on the presence or absence of an explicit nonlinear projection function between the original data space and its lower-dimensional representation.
The kernel function serves as a mechanism to simulate the projection of the original data into a high-dimensional feature space. Commonly utilised kernel functions include the polynomial kernel and the radial basis function. Several classical algorithms based on these principles include Kernel PCA, Kernel Nonnegative Matrix Factorisation (KNMF) and Polynomial Kernel Nonnegative Matrix Factorisation (PKNMF), among others. For example, Chen et al. constructed a new fractional power inner product kernel FPK, which improves the performance of the general polynomial kernel function algorithm. A novel algorithm termed ASVMK has been proposed for the elimination of Gaussian kernel bandwidth issues [112]. This approach creates a feature space intrinsic to the data's characteristics, eliminating the need for additional parameters to define neighbourhoods and thereby reducing the computation of the Gram matrix. However, during the feature mapping process, there is a risk of losing discriminative information and geometric properties inherent in the original data. Furthermore, kernel functions often suffer from a lack of interpretability, complicating the selection of an appropriate function for nonlinear transformations [83]. To address these issues, Chen et al. introduced an online adaptive kernel learning algorithm known as RF-OAK, which leverages random feature mapping to enhance the kernel function’s adaptability and responsiveness to changes in data shape [84].
Manifold learning represents another set of potent techniques for handling nonlinear data. Prominent manifold learning methods encompass Manifold Graph Construction (MGC), Locally Linear Embedding (LLE), Diffusion Maps (DM) and Maximum Variance Unfolding (MVU), etc [113]. KS-LEE is an advanced manifold learning technique that adaptively derives domain and connection weights through sparse representation, enabling the analysis of data with varied structures [85]. SNML employs an s-neighbourhood strategy for graph construction, ensuring that each pixel is linked to its eight immediate neighbours, thus preserving crucial spatial information [86]. RDLVM is an innovative method combining a weighted domain graph construction with robust discriminative latent variable manifold learning [87]. This algorithm capitalises on sample label information to bolster the cohesion of similar feature samples and the distinction between different ones, while also effectively mitigating the impact of outliers and noise. Many dimensionality reduction techniques concentrate solely on mapping from high-dimensional to low-dimensional spaces, often overlooking the validation of the low-dimensional embedding. To counter this, Yang et al. introduced a novel locality preserving projection method known as LPPAE, which is predicated on an encoder-encoder framework. This method establishes a bidirectional mapping between high-dimensional and low-dimensional spaces, thereby preserving more of the original data’s information. Consequently, the low-dimensional features can more precisely and effectively encapsulate the essence of the original high-dimensional data, enhancing the precision of feature extraction [88].
Deep Learning Methods
In addition to traditional ML for dimensionality reduction, DL has a broad spectrum of applications in feature extraction. As a data-driven approach, it constructs a hierarchical learning framework by emulating the structure of human brain neurons, thereby enhancing representation and generalisation capabilities, and enabling the extraction of deeper image features. Naeem H. and Bin-Salem A.A. developed a DL-based method for soft tissue recognition and extraction, integrating a multi-level feature extraction strategy that encompasses GIST, Scale-Invariant Feature Transform (SIFT), and CNNs to obtain dense multi-level features, where GIST and SIFT provide complementary global and local feature representations to enhance the robustness of the CNN network, significantly reducing its training complexity while improving detection accuracy [89]. Yasrab et al. designed a CNN-based encoder-decoder structure for feature extraction, combining local pixel information with global scene information to accurately segment high-resolution small root features [114]. To capture the rich spectral and spatial complementary information inherent in HSimg, He et al. proposed a novel Spectral-Spatial Aggregate Target Detector (S2ADet). This detector includes a two-stream feature extraction network that aggregates spectral and spatial information through a Spectral-Spatial Aggregation (SSA) module to extract more refined features [90]. Inspired by the two-stream architecture in behaviour recognition, Yu et al. proposed an unsupervised convolutional two-stream network (cs2GAN-FE) based on an improved Wasserstein GAN. The network employs a two-stream strategy to extract static spectral-spatial information and multi-band dynamic spectral reflectance changes simultaneously, providing a robust basis for subsequent analysis [91]. Yan et al. proposed a three-dimensional cascaded Spectral-Spatial Element Attention Network (3D-CSSEAN) to learn more meaningful features from small training sample sets of HSimg. The network integrates spectral and spatial elemental attention modules that filter data post-dimensionality reduction and extract more meaningful spectral and spatial features [92]. Zhang et al. introduced a novel finite element method based on DL. By designing a fully convolutional and deconvolutional network framework, they constructed an encoder sub-network and a decoder network, enabling end-to-end unsupervised training and reducing sample dependence [93]. These advancements showcase the versatility and potential of DL in feature extraction, offering innovative solutions to complex problems in image processing and data analysis.
Modelling
Feature extraction in the data preprocessing stage effectively removes noise and redundant information, thereby retaining only the pertinent features of the original dataset. This refined data is then suitable for subsequent modelling and analysis. Hyperspectral data modelling encompasses two primary categories: classification and regression.
Classification Modelling
Classification modelling, typically applied to images, is predominantly utilised for identifying agricultural diseases, pests, and weeds [115], as well as for terrain classification through remote sensing imagery. HSimg, with their superior spectral resolution, offer a nuanced reflection of the chemical composition and unique spectral features of subjects, alongside a wealth of spatial texture characteristics. This affords them a significant advantage in image classification and recognition tasks over RGB images [115]. In the nascent stages of HSI classification, ML-based methods predominated. Mustafa et al. leveraged HSI technology alongside multiple ML classifiers, such as SVM, KNN and RF, to investigate fusarium head blight in wheat [94]. Addressing the challenge of limited datasets, Makantasis introduced a novel ML model grounded in tensor algebra operations, integrating a learning algorithm to refine the classifier. For nonlinear classification issues, a model was introduced with weights that adhere to a rank-1 regular decomposition property [116]. Zhang et al. proposed an unsupervised HSI classification approach based on Robust PCA to circumvent issues arising from a paucity of labelled samples [95]. Despite these methods exhibiting commendable classification efficacy, the inherent complexity and spectral information interdependencies and redundancies in HSI data processing render traditional methods ill-suited for the intricate and variable scenarios.
Data-driven DL offers a more dynamic feature extraction process and demonstrates enhanced flexibility in image classification applications. Zhang et al. proposed a Deep Residual CNN (DRCNN) model that amalgamates various pixel regions to glean richer information and incorporates multi-scale summation modules for feature aggregation, thereby enhancing classification precision [96]. The proposed stochastic window transformer (SWFormer) framework enhances hyperspectral-LiDAR classification by leveraging parallel spatial-spectral feature extraction, multi-scale strip convolution, and a SWFormer with adaptive feature aggregation, effectively capturing discriminative local-global features while reducing computational overhead for improved accuracy [117]. Wang et al. proposed a CNN-based classification method, CVNN, which embeds vector neuron capsule representations into a Fully Convolutional Network (FCN) framework to bolster classification accuracy through enhanced classification steps, particularly in scenarios with insufficient sample labels, and further addressed the class imbalance issue in HSimg datasets by combining ML with DL techniques, including oversampling via the SMOTE algorithm followed by CNN-based classification [97]. In scenarios where obtaining or labelling HSimg is challenging, Liu et al., building upon domain adaptation methodologies and matrix factorisation-based data clustering research, introduced a network, CDA, capable of simultaneous feature extraction and classification. This network leverages adversarial adaptation and probabilistic maximum average discrepancy, eschewing the need for target labels and enabling unsupervised classification [118]. Drawing inspiration from the Siamese architecture, which operates unsupervised and learns the intrinsic structure of samples, Hao et al. proposed a deep network architecture for HSimg classification, SRCL, that is assisted by a super-resolution network. This architecture, with the aid of transfer learning and unsupervised training, effectively addresses data scarcity and adaptively learns specific classification tasks [98].
Regression Modelling
Hyperspectral data analysis encompasses not only classification but also quantitative regression, which is equally significant. Several studies have effectively employed ML regression models for hyperspectral data analysis. For instance, Flynn et al. developed a model using in situ hyperspectral data to predict nitrogen concentration, biomass, and nitrogen content through KNN, PLS, SVM and RF ML techniques [99]. Furlanetto et al. investigated the relationship between hyperspectral reflectance and potassium content in soybean leaves, employing various ML regression models [25]. Despite the recognition of ML methods in estimating material content, challenges persist due to the unique characteristics of hyperspectral data, such as noise and multicollinearity, which can result in overfitting and misinterpretation. Tailored algorithms, learning strategies, and predictors are essential for different scenarios. Building on MLR, Fei et al. proposed a feature splitting regression (FSR) model to estimate the yield and other traits of winter wheat, utilising a multidimensional feature segmentation strategy that enhances model robustness and applicability [119]. Guo et al. evaluated the performance of multiple ML algorithms for predicting corn grain yield using UAV-based hyperspectral imagery. Among the tested models, RF and LightGBM demonstrated superior accuracy when applied to vegetation indices (VIs) derived from sensitive spectral wavelengths, achieving R2 values of 0.90 (RMSE = 0.55 t/ha) and 0.85 (RMSE = 0.59 t/ha), respectively. Notably, the incorporation of full hyperspectral spectra further enhanced the prediction accuracy, with RF attaining the highest performance (R2 = 0.92, RMSE = 0.53 t/ha) [100].
DL, an emerging subfield of ML, has also found applications in quantitative regression. Zhang et al. explored the use of convolutional neural networks and deep autoencoders for predicting nutrient elements in black wolfberry, achieving promising results [101]. Rehman et al. developed an end-to-end DL model, DeepRWC, based on one-dimensional convolutional neural networks, which can predict the relative water content of plants from average spectral reflectance with higher accuracy than PLSR and SVM [102]. Deng et al. employed DL semantic segmentation for pixel-level image regression, providing insights for high-throughput crop phenotypic analysis, pest detection, and yield assessment [120]. Integrating information from multiple dimensions to extract more comprehensive features is a crucial research direction in hyperspectral data processing. Liu et al. combined hyperspectral data with time series phenotypes to predict SSC, pH value, nitrate, calcium, and water stress in lettuce, using a DL model [121]. A novel method for fusing and extracting temporal-spatial-spectral information was established, combined with a CNN for predicting soil organic carbon content, which showed higher accuracy than models relying on simple relationships [103].
As technology advances, innovative methods for hyperspectral data analysis continue to emerge. The challenge of extracting target features more completely, accurately, and rapidly from multi-domain, small-sample scenarios remains a focal point in the field. Further research is necessary to expand the application of HSI technology across various sectors.
Next-Gen Data Processing and Cross-Domain Applications
Hyperspectral data analysis is progressing through AI-based approaches that address several technical challenges: (1) Automated processing using end-to-end pipelines that incorporate noise reduction, data compression, and feature extraction; (2) Advanced modelling using self-supervised learning and generative AI to overcome data limitations while maintaining interpretability via attention mechanisms; and (3) Integration with complementary data sources such as LiDAR and thermal imaging to enhance environmental analysis. These developments are improving the transformation of hyperspectral data into practical applications.
Rising Tides of Innovation: Emerging Trends in HSI
Deep Learning Methods
In recent years, advancements in information technologies, particularly artificial intelligence, have led to the application of numerous innovative techniques and methods in HSI. As mentioned above, DL methods have been extensively applied to HSI for tasks such as noise reduction, dimensionality reduction and data modelling. Notably, three-dimensional convolutional neural networks (3D-CNNs) can simultaneously extract spectral and spatial features. For instance, Díaz-Martínez et al. employed a full-spectral hypercube of rice seeds to train a 3D-CNN for classifying seeds exposed to varying temperature conditions, with the model showing reduced performance at the highest day-night temperature (36°C–32°C), confirming the impact of thermal stress on spectral discriminability [122]. Ji et al. demonstrated that a 3D-CNN, which convolves spatial and temporal dimensions by integrating the strengths of 1D-CNN and 2D-CNNs, achieved significantly better crop classification performance than 2D-CNN, further highlighting the importance of temporal features [123]. Zhao et al. proposed an HSI-based method combining 3DCNN and LSTM to extract joint spectral-spatial features for robust in situ LCC estimation, achieving high accuracy (R2 up to 0.96) across multiple years and growth stages while reducing soil interference and segmentation dependency [124]. Wang et al. designed a 3D denoising convolutional neural network (3DDnCNN) to recover clear images with multiple spectral channels from blurred and noisy observations, emphasising the use of the Alternating Direction Method of Multipliers (ADMM) to decompose the optimisation problem [125]. Additionally, 3D-CNNs have been employed in Generative Adversarial Networks (GANs). Xue et al. proposed 3D-CNN-GAN framework that addresses limited labelled samples in HSimg classification by generating synthetic spectral-spatial cubes through adversarial training between a 3D-CNN generator and 3D residual discriminator, enhancing classification performance with scarce annotations [126]. In the field of HSimg super-resolution (SR), Dou et al. introduced a novel 3D attention-based SRGAN (3DASRGAN) that enhances spectral information in SR by utilising 3D convolution based on the SRGAN structure and incorporating an attention mechanism for managing multiple features [127].
The amalgamation of transfer learning with GAN capitalises on their data generation prowess, enabling cross-domain knowledge transfer and enhancing performance and generalisability across diverse tasks. This model transfer approach adapts spectral data or calibration models to various conditions, circumventing the need for extensive recalibration. Padarian et al. fine-tuned a model pre-trained on a neural network with a modest amount of local data from the LUCAS database, achieving a 91% reduction in error rates and highlighting the efficacy of transfer learning in soil spectroscopy [128]. Rehman and Jin used adversarial learning to transfer DL models across different imaging facilities, employing Domain Adversarial Neural Network (DANN) and Adversarial Discriminative Domain Adaptation (ADDA) to address domain shift and produce more domain-specific features [129]. Besides, Ye et al. proposed a multi-cycle adversarial transfer learning framework (CDMC) for cross-domain HSimg classification, extending CycleGAN with iterative error-accumulated mappings and auxiliary classifiers to address sensor-induced domain shifts under limited target labels [130]. In essence, the fusion of transfer learning with GANs emerges as a robust strategy for calibrating spectral data and models to a multitude of scenarios, markedly augmenting the precision and applicability of HSI.
Reconstructing HSimg helps to fully utilise existing data, reduce noise and improve data quality. Xu et al. proposed a region-based block compressed sensing (RBCS) algorithm for reconstructing HSimg, employing local mean and local standard deviation (LMLSD) criteria to select the optimal band and introducing k-means clustering to extract regions from the optimal band [131]. To address the Hughes phenomenon and high computational costs due to continuous narrow band correlation in HSimg, the Band Selection Network (BS Net) was proposed [132]. This network uses the Band Attention Module (BAM) to model nonlinear dependencies between spectral bands and includes a Reconstruction Network (RecNet) to restore the original image from information-rich bands. The SSR-NET network integrates low-spatial-resolution HSimg with high-spatial-resolution multispectral images to reconstruct high-spatial-resolution HSimg, starting with CMMI to generate fused images, followed by spatial and spectral reconstruction to recover lost information [133]. These advancements demonstrate the significant impact of DL on enhancing the precision and applicability of HSI.
New Applications of Integrating HSI With Other Optical Imaging Techniques
In recent years, the integration of HSI with other cutting-edge techniques has revolutionised the field of biological and agricultural research. This synergy has not only enhanced the capabilities of HSI but also contributed to more precise and efficient analytical outcomes.
One notable application is the combination of HSI with microscopic techniques (HSI-microscopy), which facilitates the detection of minute changes within plant tissues at the cellular level. For instance, Kuska et al. utilised a hyperspectral microscope to study the spectral changes in barley leaves during resistance reactions against powdery mildew. This approach enabled the differentiation of barley genotypes based on their susceptibility to the disease, demonstrating the potential of HSI for automated characterisation of plant-pathogen interactions [34]. Expanding on this, Zhang et al. demonstrated the innovative potential of HSI-microscopy for cellular-level phenotyping by developing a Fusarium head blight (FHB) classification framework. Their workflow leveraged hyperspectral microscopy images of winter wheat spikelets to construct a novel classification index, integrating the instability index (ISI) with established analytical methods—including the spectral angle mapper (SAM) for wavelength selection and PLSR for spectral index optimisation. This HSI-microscopy system achieved 89.80% classification accuracy, highlighting its effectiveness for early FHB detection and underscoring the broader utility of combining HSI with microscopic techniques in precision agriculture applications [134]. Furthermore, Hao et al. developed an innovative approach integrating micro-HSI (MHSI) and Vis-NIR HSI with a hybrid CNN-LSTM DL architecture and heterogeneous two-dimensional correlation spectroscopy (H2D-COS) for accurate prediction of superoxide dismutase (SOD) activity in tomato leaves under stress conditions, achieving exceptional prediction performance (RP = 0.9075 for microscopic-scale analysis) while establishing effective cross-scale transfer learning capabilities (RP = 0.7549), thereby demonstrating the significant potential of multimodal HSI systems for precise physiological monitoring in plant stress responses [135].
The convergence of HSI with Raman spectroscopy has revolutionised the field of molecular imaging and chemical analysis, providing unprecedented insights into the composition and structure of biological and chemical systems. Karpf et al. introduced a time-encoded technique for fibre-based hyperspectral broadband stimulated Raman microscopy. By harnessing continuous wave, rapidly swept probe lasers alongside a short-duty-cycle, actively modulated pump laser, they developed a system that delivers high-speed operation, extensive spectral coverage, and high resolution. This breakthrough enables precise chemical analysis and hyperspectral Raman microscopy, offering molecular contrast that is invaluable for biological and chemical research [136]. In the realm of label-free imaging, HSI has been teamed with such techniques to scrutinise the localisation of specific compounds within plant tissues. Long et al. utilised label-free Raman HSI (358–1700 cm−1) to nondestructively classify corn kernel viability during ageing, achieving > 97% accuracy through whale optimisation algorithm (WOA)-optimised SVM/ELM models coupled with feature selection algorithms. By integrating spectral and spatial resolution, this approach identified ageing-related spectral signatures (e.g., peaks at 475 cm−1 for starch) without requiring destructive sampling, which demonstrates Raman HSI's potential to replace conventional viability assays (e.g., TTC staining, germination tests) for in situ metabolic profiling in agricultural products [137].
Hyperspectral fluorescence imaging is increasingly being utilised due to its capability to capture simultaneous spatio-temporal dynamics at various scales, ranging from molecules to cells and tissues, using multiple fluorescent markers. In a related study, Zou et al. employed fluorescence hyperspectral imaging (FHSI) covering 400–1000 nm spectral range at 2.8 nm resolution, utilising 357 nm excitation and 475–575 nm emission filters to capture 125 spectral channels per kiwifruit sample. With SG-MSC preprocessing, the developed improved WOA-CNN-MLP model achieved 91.6% classification accuracy for kiwifruit maturity stages while maintaining a low loss value of 0.23 [138]. Zhou et al. have harnessed the capabilities of fluorescence HSI combined with DL algorithms for lead (Pb) detection in oilseed rape leaves. They have successfully integrated wavelet transform (WT) with a stacked denoising autoencoder (SDAE) to extract sophisticated features [107]. Moreover, to amplify the precision of Pb detection across both silicon-free and silicon-rich environments, they proposed an innovative transfer stacked convolution auto-encoder (T-SCAE) algorithm [139]. Wang et al. demonstrated the effectiveness of the fluorescence HSI system in detecting AFB1 contamination and provided a theoretical basis for developing new detection and estimation methods for the corn industry [140]. And Chun et al. explored the use of hyperspectral fluorescence imaging for the early detection of Botrytis cinerea infection in strawberry fruits. This study develops a 1D-CNN model based on partial least squares-discriminant analysis (PLS-DA), VGG-19, and ResNet-50. The application of data augmentation techniques and spectral preprocessing significantly improves the model’s performance, with the ResNet-50-based 1D-CNN model showing the highest accuracy [141].
Finally, video spectral imaging, essential for remote sensing, facilitates the acquisition of 4D information (2D spatial dimensions, spectrum and time), which is highly beneficial for the detection of dynamic targets. Xu et al. developed a groundbreaking ‘Single-Doxel Imager’ (SDI) system using a single-pixel detector and compressive sensing (CS) principles. This system achieves video-rate HSI (4.3 fps) at a remarkable 900× data compression ratio, reconstructing 128 × 128-pixel scenes with 64 spectral bands. By jointly encoding spatial-spectral information via a digital micromirror device (DMD) and exploiting 4D sparsity with optical flow-assisted 4D total variation regularisation, the SDI overcomes bandwidth limitations while maintaining a high signal-to-noise ratio [142]. In summary, the amalgamation of HSI with a suite of other technologies has propelled its evolution as a multifaceted tool for scientific inquiry. Whether in the context of plant pathology, precision agriculture, molecular biology or cellular imaging, these integrated approaches have expanded the horizons of what is discernible and achievable, paving the way for future breakthroughs in research and application.
Portable Hyperspectral Imaging Devices
Portable and handheld HSI devices provide enhanced portability and can be used for field measurements without being limited by the laboratory environment, thereby enhancing flexibility and ease of use. In addition, they are generally proven to be more cost-effective compared to the HSI systems used in the laboratory.
The LeafSpec device, developed at Purdue University, is a portable hyperspectral imager designed for accurate and low-cost crop leaf phenotyping. It has been successfully tested in both field and greenhouse conditions, effectively detecting differences in nitrogen treatments and genotypes in corn plants [143]. The FieldSpec HandHeld 2 portable spectroradiometer (ASD Inc.) was used to measure canopy-level hyperspectral reflectance (325–1075 nm) of rapeseed, providing high-resolution spectral data that improved the accuracy of UAV-based vegetation indices for crop growth monitoring [144]. Similarly, using ASD Handheld 2, canopy reflectance data of tea plants were collected and fused with wavelet features and ML (RF and LASSO), enabling accurate year-round monitoring of biomass and nitrogen accumulation [145]. Recently, the HAIP Black Mobile spectral imaging sensor, a cutting-edge mobile hyperspectral device designed for rapid, non-destructive data acquisition in agricultural environments, was developed by HAIP Solutions (). The sensor's portability, real-time imaging capabilities, and compatibility with ML pipelines position it as a promising tool for high-throughput phenotyping and on-site decision-making in precision agriculture. Its accuracy under different field conditions and crop species requires further validation.
Handheld HSI devices also have significant applications in plant growth analysis. For example, the Specim IQ captured 420 rice images under 14 NPK stress conditions, enabling spectral analysis and the development of the SHCFTT model, which achieved 93.92%–100% accuracy in nutrient stress identification [24]. In remote sensing, a high-performance HSI spectrometer named HSIS-SIF has been developed for observing Solar-Induced Chlorophyll Fluorescence (SIF) in vegetation [146]. The Scanning Plant IoT (SPOT) facility, a laboratory-based platform integrating a hyperspectral sensor, thermal camera, and LiDAR camera, was developed at the University of Florida to collect high-quality plant phenotypic data [147]. The open-source nature of SPOT facilitates hypothesis-driven research and bridges the gap between field-scale UAV data and laboratory experiments. Finally, the OpenHSI project represents a significant step towards making HSI technology more accessible. By developing an open-source optical design and a comprehensive software platform, it provides a complete solution for capturing, calibrating and processing hyperspectral data [148]. The successful field testing and verification of the system's performance underscore its potential for diverse applications.
Overall, portable and handheld HSI devices have become indispensable tools for a wide range of agricultural and environmental applications. Their portability, cost-effectiveness and capacity to deliver real-time data render them highly suitable for field-based measurements and informed decision-making in precision agriculture. With ongoing advancements in technology, these devices are poised to significantly enhance our comprehension of plant health, soil characteristics, and various other environmental parameters. This enhanced understanding will, in turn, contribute to the development of more sustainable and efficient agricultural practices.
Innovations and Future Directions
Recent advancements in DL, integration with complementary optical techniques, and the development of portable devices are expanding the use of HSI in biological and agricultural applications. Future progress in this field may focus on: (1) Optimising lightweight DL models for efficient edge computing, enabling real-time HIS analysis in field conditions; (2) Developing HSI systems for root imaging to improve understanding of plant root systems; (3) Enhancing multimodal data fusion frameworks to integrate HSI with other imaging modalities for more comprehensive analysis; (4) Improving portable HSI devices to increase resistance to environmental interference (such as ambient light) and simplify operation for practical applications.
Toolkits of Hyperspectral Data Processing
In recent years, DL technologies have profoundly transformed the processing of hyperspectral data. Efficient computational resources have been employed to manage vast amounts of hyperspectral data. Consequently, the field of hyperspectral data processing has exhibited a diversified technological landscape. Commercial software such as ENVI and MATLAB are widely adopted in this domain, while integrated toolkits like HSI-PP have gained considerable recognition [149]. Additionally, a plethora of custom-coded solutions, leveraging programming languages such as Python, C, C++ and MATLAB [150], are favoured by scholars. To facilitate researchers, numerous targeted plugins for tools like MATLAB and ENVI have emerged, alongside scholars utilising spectroradiometers to directly acquire sample reflectance for subsequent operations [99]. The detailed distribution of hyperspectral data processing approaches across various programming languages is illustrated in Figure 4a, indicating Python’s highest frequency of use in hyperspectral data processing, reaching 46%, closely followed by MATLAB and C/C++.
[IMAGE OMITTED. SEE PDF]
Hyperspectral data find widespread applications in various domains, including classification, denoising, image fusion and spectral reconstruction, among others. Classification tasks, in particular, occupy a considerable proportion, as depicted in Figure 4b, detailing the distribution across different domains. The functional characteristics, primary programming languages, and the location of relevant code resources in the field of hyperspectral data processing in recent years are summarised in Table 2, providing significant scholarly reference for relevant researchers.
TABLE 2 Software or code links corresponding to hyperspectral data processing in different application scenarios.
| Function | URL |
| ENVI (plugin) | |
| Software | |
| Segmentation | |
| Feature selection | |
| Classification | |
| Unmixing | |
| Denoising | |
| Anomaly detection | |
| Fusion | |
| Reconstruction | |
| Clustering | |
| Other | |
This chapter provides a comprehensive overview of hyperspectral data processing software and code resources, systematically organised by target system, offering researchers a centralised reference and access point. Future developments in this domain could focus on: (1) The creation of an intelligent platform that provides customised code solutions based on user-described requirements, integrating research papers, executable code and associated datasets to improve reproducibility; and (2) The adoption of a unified open-source framework (e.g., a Python library with Docker support) to standardise data preprocessing and ML workflows. Additionally, establishing standardised hyperspectral data repositories with mandatory metadata (e.g., spectral resolution, acquisition conditions) and API-based access could enhance data consistency and interoperability across studies.
Hyperspectral Databases
Since the application of HSI in plants, a large number of databases have been generated, including the raw binary database and image database, reflectance database and hyperspectral vegetation indices database, which have laid a solid foundation for smart agriculture research.
Raw Binary Database and Image Database
During the 1990s, in tandem with the advancements in hyperspectral remote sensing technology, researchers worldwide undertook numerous spectral measurements, resulting in the acquisition of a wealth of spectral data and the establishment of extensive hyperspectral databases. Typically, the collected HSimg binary data would be stored using formats such as band-interleaved-by-line (BIL) files, band sequential (BSQ) files and others [151]. The opening of these file formats can be facilitated using ENVI software. The raw binary data can be reorganised to obtain the corresponding HSimg. Consequently, when distributing the HSimg database, certain researchers opted to provide the accompanying raw binary data files. For example, the WHU–Hi dataset is showcased, encompassing the unveiling of UAV HSimg (in Tiff format) and the associated original binary data files (in ENVI standard format) for diverse crop types encountered in Hubei Province, China [152]. Despite the growing availability of HSI data, open access to raw spectral data remains scarce. For example, Bohnenkamp et al. developed a comprehensive hyperspectral library documenting spectral progression in five wheat foliar fungal diseases (e.g., brown rust, powdery mildew) from infection to symptom manifestation, achieving remarkable detection accuracy (≤ 99%) through ML classifiers [153]. Regrettably, this valuable resource has not yet been made publicly accessible, limiting its broader utilisation for research and agricultural applications. Here, publicly available hyperspectral databases in plant research are systematically catalogued in this study, as summarised in Table 3. Analysis of the 26 studies listed revealed that raw binary data were accessible in 11 publications, while HSimg data were openly shared in 20 cases. Notably, only 10 studies provided both data types concurrently.
TABLE 3 List of the public hyperspectral databases containing raw binary data, image and reflectance data information.
| Type | Species/plant | Database type | Data description | Reference | ||
| Raw binary data | Image | Reflectance data | ||||
| Horticulture | Olive | √ | √ | √ | An in-field HSimg sensor was used for olive data acquisition during the table-olive season from May to September. | [154] |
| Grape | √ | √ | √ | Hyperspectral data (range from 400 to 1000 nm) of 3 different grape varieties were collected in Gaillac (France), in summer 2020. | [155] | |
| 205 grape plant leaves were measured with a hyperspectral camera in the visible/near infrared spectral range under controlled conditions. | [156] | |||||
| Potato | √ | √ | — | The hyperspectral imagery of potato fields was collected using a push-broom line imaging spectrometer with 150 spectral bands in the range of 460–902 nm, that was mounted on a UAV. | [157] | |
| Fruit and vegetable | √ | — | — | Samples of apple, broccoli, leek, and mushroom, were measured by hyperspectral cameras in the visible/near-infrared spectral domain (430–900 nm). | [158] | |
| Apple | √ | √ | — | The data presents a temporal monitoring of apple tree plants, infected or not with fire blight, in a sequence of HSimg. | [154] | |
| — | √ | — | The dataset presents two series of HSimg of healthy and infected apple tree leaves acquired daily, from 2 days after apple scab disease inoculation until an advanced stage of infection (11 days after inoculation). | |||
| The dataset provides three classes of HSimg: Pure, insecticide-immersed, and fungicide-immersed apples with different concentrations of fertilisers. | ||||||
| Flower | — | √ | — | The dataset named HFD100 contains more than 10,700 HSimg which belong to 100 categories for flower classification. | [159] | |
| Spearmint | — | — | √ | The reflectance spectral data (935–1720 nm) of mint samples and corresponding essential oil data. | [160] | |
| Tomato | — | — | √ | The dataset included hyperspectral data of plants of two cultivars of Solanum lycopersicum: Benito and Polfast, which were infected with five different pathogens. | [161] | |
| Crop and horticulture | Crops and trees | √ | √ | √ | A distinct dataset of high-resolution hyperspectral imagery and associated ground truth spectra of various vegetable crops acquired over a tropical forest ecosystem. | [162] |
| — | — | √ | 1394 leaves reflectance corresponding to 60 different annual and perennial plant species, which grew in forests, wild, and as domesticated crops under various developmental stages and growing conditions. | [163] | ||
| Corn and peanut | √ | √ | √ | The HyperPRI dataset features HSimg of corn and peanut plant roots in rhizoboxes, annotated for root and soil analysis, useful for studying root traits and drought responses. | [164] | |
| Variety of plants | — | √ | √ | The dataset is consisted of three group files: (1) the figure and hyperspectral reflectance data of 10 different land surface types, including Masson pine, bamboo forest, tea plant, reed, paddy, sweet potato, caraway, weed, water body and building/road; (2) the hyperspectral remote sensing reflectance image over Fanglu tea farm; (3) the ground truth data in the Fanglu tea farm. | [165] | |
| — | √ | — | The dataset features an aerial HSimg of Xiongan new area, taken in October 2017. It includes 19 labelled land cover types, predominantly cash crops. | [166] | ||
| — | √ | — | Various HSimg datasets captured by sensors like AVIRIS and ROSIS. These images depict diverse scenes such as crops, forests, and urban areas, with detailed ground truth data for classification and analysis purposes. | |||
| Crop | Corn | √ | √ | — | The raw hyperspectral data and images of corn under different drought severities and the third part of the data collected during the transferability (TF) experiment. | [167] |
| — | √ | — | The image data collected from 155 corn plants over 32 days representing 32 corn inbreds. Each imaging time point included data from four different types of cameras: RGB, hyperspectral, fluorescent and thermal IR. | [168] | ||
| A simple dataset comprising three distinct types of local corn seeds in Ghana. The dataset is presented in two parts: Raw images, consisting of 4846 images, are categorised into bad and good. Augmented images consist of 28,910 images, with 13,250 representing bad data and 15,660 representing good data. | [169] | |||||
| — | — | √ | The hyperspectral reflectance data of 368 corn genotypes plants with or without drought stress over a course of 98 days (Day 34, 40, 46, 52). | [45] | ||
| Wheat | √ | √ | — | The dataset contains HSimg (.dat format) of four wheat lines, each with a control and a salt (NaCl) treatment. | [170] | |
| — | √ | — | A meticulously curated and annotated dataset, named as SPIKE-segm, taken from the publicly accessible SPIKE dataset. | [171] | ||
| Rice | — | — | √ | 1540 hyperspectral indices at whole-plant level during tillering, heading, and ripening stages of 529 rice accessions. | [172] | |
| The reflectance data were collected using a handheld leaf spectrometer on rice plants in the seedling stage. | [173] | |||||
| Soybean | — | — | √ | Leaf reflectance data from 2012 to 2013 in the SoyNAM population. | [174] | |
| Brassica napus | — | — | √ | The hyperspectral traits, including the total reflectance-related traits, average reflectance-related traits and logarithm-related traits of 505 B. napus accessions under different salt stress conditions. | [44] | |
| Arabidopsis thaliana | — | — | √ | Leaf hyperspectral reflectance data for 721 widely distributed A. thaliana natural accessions grown in a common garden experiment. | [175] | |
| Variety of crops | √ | √ | — | The WHU–Hi dataset is made up of the WHU–Hi–LongKou, WHU–Hi–HanChuan and WHU–Hi–HongHu datasets. All the datasets were acquired in farming areas with various crop types in Hubei province, China, and WHU–Hi–HongHu dataset contains up to 18 types of crops. | [152] |
Reflectance Database
Reflectance data stands out as the primary and most extensively utilised dataset in HSI technology. The initial hyperspectral reflectance databases accessible to the public encompassed the USGS, ASTER, and the vegetation spectral library at Johns Hopkins University. Subsequently, these databases were incorporated into commercial software platforms such as ENVI, PCI and ERDAS, which providing the essential groundwork for HSimg data processing. Over the past approximately 25 years, the public availability of hyperspectral reflectance data has been relatively scarce [176]. Traditionally, researchers mainly relied on obtaining this data directly from the corresponding author. More recently, with contributions from various institutions, an open-access Ecological Spectral Information System (EcoSIS, ) spectral database has been established, which encompasses leaf and canopy spectra of trees, shrubs, grasses, crops, etc., in recent decades. Nevertheless, in recent times, fuelled by the emergence of big data and the imperative to accelerate the development of new crop varieties and implement precise field management strategies, there has been a gradual transition towards increased data transparency. While the majority of the articles do not furnish the outcome of the reflectance data, the associated spectral reflectance curves are predominantly revealed (Figure 5).
[IMAGE OMITTED. SEE PDF]
Hyperspectral Vegetation Indices Database
HVIs not only have the potential to identify but also quantify variations in plant status, such as nutrient levels, which are essential for effective dynamic management and forecasting of crop productivity [178]. Over decades of research efforts, a comprehensive database of HVIs has been established, encompassing a wide range of indices, as shown in Table 4. This database includes structural indices such as NDVI, RDVI (Renormalised Difference Vegetation Index) and OSAVI (Optimised Soil-Adjusted Vegetation Index). It also features nitrogen indices (NDDA, mSR705, NDNI), water indices (NDWI, R970/R900), chlorophyll indices (MCARI, TCARI, TVI), carotenoid indices (PRI, Cx + c), anthocyanin indices (ARI), cellulose indices (CAI), protein indices (NDSI) and health indices (CLS, SBRI).
TABLE 4 List of hyperspectral vegetation indices (HVIs).
| Index type | Vegetation index | Equation | Reference | |
| Structural indices | SIPI-1 | Structure insensitive pigment index-1 | (R800 − R445)/(R800 − R680) | [179] |
| SIPI-2 | Structure insensitive pigment index-2 | (R800 − R435)/(R415 − R435) | ||
| GI | Green index | R554/R677 | ||
| EVI | Enhanced vegetation index | 2.5 × [(R900 − R680)/(R900 + 6 × R680 − 7.5 × R475 + 1)] | ||
| EVI800, 660 | Enhanced vegetation index | 2.56 × (R800 − R660)/(1 + R800 + 2.4 × R660) | ||
| TGI | Triangular greenness index | −0.5 × [190 × (R670 − R550) − 120 × (R670 − R480)] | [23] | |
| NDVI | Normalised difference vegetation index | (R800 − R670)/(R800 + R670) | [179] | |
| RDVI | Renormalised difference vegetation index | (800 − R670)/(R800 + R670)1/2 | ||
| OSAVI | Optimised soil-adjusted vegetation index | (1 + 0.16) × (R800 − R670)/(R800 + R670 + 0.16) | ||
| MCARI1 | Modified chlorophyll absorption in reflectance index 1 | 1.2 × [2.5 × (R800 − R670) − 1.3 × (R800 − R550)] | ||
| MSAVI | Modified soil adjusted vegetation index | 1/2 [2 × R800 + 1 − ((2 × R800 + 1) − 8 × (R800 − R670))1/2] | ||
| GNDVI | Green normalised difference vegetation index | (R790 − R550)/(R790 + R550) | [180] | |
| NDRE | Normalised difference red edge index | (R790 − R720)/(R790 + R720) | ||
| RVI | Ratio vegetation index | R790/R670 | ||
| Nitrogen indices | DCNI | Double-peak canopy nitrogen index | (R720 − R700)/(R700 − R670)/(R720 − R670 + 0.03) | [181] |
| BNI1 | Blue nitrogen index 1 | R434/(R496 + R401) | ||
| BNI2 | Blue nitrogen index 2 | (R498 + R413)/R442 | ||
| NDDA | Normalised difference index of the double-peak areas | (R755 + R680−2 × R705)/(R755 − R680) | ||
| mSR705 | Modified red-edge ratio | (R750 − R445)/(R705 − R445) | ||
| mRER | Modified red-edge ratio | (R759 − 1.8 × R419)/(R742 − 1.8 × R419) | ||
| NDNI | Normalised difference nitrogen index | [log(1/R1510) − log(1/R1690)]/[log(1/R1510) + log(1/R1690)] | [182] | |
| Water indices | EWT | Equivalent water thickness | (R1390 − R1370)/(R1390 + R1370) | [183] |
| RWC | Relative water content | (R1620 − R1410)/(R1620 + R1410), (R2160 − R2090)/(R2160 + R2090) | ||
| NDWI | Normalised difference water index | (R860 − R1240)/(R860 + R1240) | [184] | |
| NWI970, 990 | Normalised water index (R970, R990) | (R970 − R900)/(R970 + R900) | ||
| R970/R900 | — | R970/R900 | [185] | |
| R1300/R1450 | — | R1300/R1450 | ||
| Chlorophyll indices | PBI | Plant biochemical index | R810/R560 | [186] |
| VREI1 | Vogelmann red-edge index 1 | R740/R720 | ||
| VREI2 | Vogelmann red-edge index 2 | (R734 − R747)/(R715 + R726) | ||
| AutoVI-chl | Automated hyperspectral vegetation index chlorophyll index | [(R1607 − R1384) − (R1607 − R716) × (R1607/R1384)]/[2 × ((R610 − R1384)/(R610 + R1384 + 1))] | [187] | |
| dND735, 573 | Normalised differences type index using reflectance derivatives at 735 and 573 nm | (D735 − D573)/(D735 + D573) | [188] | |
| NPCI | Normalised pigments chlorophyll ratio index | (R680 − R430)/(R680 + R430) | [189] | |
| MCARI705, 750 | Modified chlorophyll absorption ratio index calculated with reflectance from 705 to 750 nm | [(R750 − R705) − 0.2 × (R750 − R550)] × (R750/R705) | ||
| MCARI2 | Modified chlorophyll absorption ratio index improved | 1.5 × [2.5 × (R803 − R671) − 1.3 × (R803 − R549)]/[(2 × R803 + 1)2 − (6 × R803 − 5 × R6711/2) − 0.5]1/2 | ||
| MCARI1 | Modified chlorophyll absorption ratio index 1 | 1.2 × [2.5 × (R800 − R670) − 1.3 × (R800 − R550)] | [190] | |
| TCARI | Transformed chlorophyll absorption in reflectance index | 3 × [(R700 − R670) − 0.2 × (R700 − R550) × (R700/R670)] | ||
| TCARI/OSAVI | Transformed chlorophyll absorption in reflectance Index/Soil-adjusted index | 3 × [(R700 − R670) − 0.2 × (R700 − R550) × R700/R670)]/[(1 + 0.16) × (R800 − R670)/(R800 + R670 + 0.16)] | ||
| TVI | Triangular vegetation index | 0.5 × [120 × (R750 − R550) − 200 × (R670 − R550)] | ||
| G | — | R550/R670 | ||
| ZTM | Zarco–Tejada and Miller index | R750/R710 | ||
| VOG | Vogelmann index | R740/R720 | ||
| CIred edge | Red edge chlorophyll index | R790/R720 − 1 | [180] | |
| CIgreen | Green chlorophyll index | R790/R550 − 1 | ||
| MTCI | MERIS terrestrial chlorophyll index | (R750 − R710)/(R710 − R680) | ||
| Carotenoid indices | CARI | Carotenoid index | (R720/R510) − 1 | [190] |
| CRI | Carotenoid reflectance index | 1/R510 − 1/R550, 1/R510 − 1/R700 | ||
| dND516, 744 | Normalised differences type index using reflectance derivatives at 516 and 744nm | (D516 − D744)/(D516 + D744) | [191] | |
| Cx + c | Carotenoid index | R515/R570 | [192] | |
| LIC3 | — | R440/R740 | ||
| RARS | Related to the carotenoids content, as well as chlorophyll a + b | R746/R513 | ||
| PRI | Photochemical reflectance index | (R570 − R531)/(R570 + R531) | ||
| PRI515 | Photochemical reflectance index (515) | (R515 − R531)/(R515 + R531) | ||
| PRIn | Photochemical reflectance index | PRI/(RDVI × R700/R670) | ||
| PRI*CI | Carotenoid/chlorophyll ratio index | (R680 − R500)/R750 | [193] | |
| PSSR | Pigment-specific simple ratio | R800/R680, R800/R635, R800/R470 | ||
| PSND | Pigment-specific normalised difference | (R800 − R470)/(R800 + R470) | ||
| Anthocyanin indices | ARI1 | Anthocyanin reflectance index l | 1/R550 − 1/R700 | [190] |
| ARI2 | Anthocyanin reflectance index 2 | R800 × (1/R550 − 1/R700) | ||
| Cellulose index | CAI | Cellulose absorption index | 0.5 × (R2000 + R2200) − R2100 | [194] |
| Protein index | NDSI | Normalised difference spectral index | (R1227 − R1177)/(R1227 + R1177) | [195] |
| ‘Healthy’ indices | HI | Healthy index | (R534 − R698)/(R534 + R698) − 1/2 × R704 | [196] |
| CLS | Cercospora leaf spot index | (R698 − R570)/(R698 + R570) − R734 | ||
| SBRI | Sugar beet rust index | (R570 − R513)/(R570 + R513) + 1/2 × R704 | ||
| PMI | Powdery mildew index | (R520 − R584)/(R520 + R584) + R724 |
Building Comprehensive Hyperspectral Libraries
This chapter systematically catalogues available public resources, including raw data, images, reflectance, and HVIs. Despite their significant value for plant science and agricultural applications, the current scope of openly accessible, comprehensive hyperspectral databases remains limited relative to the expanding demands of HSI research. To bridge this critical gap, future initiatives must prioritise: (1) Increasing open access to raw hyperspectral data—particularly for staple and economically important crops—in standardised formats compatible with common analytical platforms (e.g., ENVI, MATLAB); and (2) Developing structured spectral datasets that encompass key variations across climate zones, genetic cultivars, and biotic stressors (e.g., specific pathogens and insect pests). The creation of well-curated, large-scale spectral libraries would support the development of reliable hyperspectral processing workflows and ML models, facilitating broader adoption of HSI technology for agricultural monitoring and management.
Discussion and Future Perspectives
Exploration of Novel Hyperspectral Instrument
In traditional imaging spectrometers, improving spectral resolution often comes at the expense of spatial and temporal resolution, resulting in slower data acquisition speeds. To address this issue, CS technology has emerged. By employing sparse sampling and reconstruction algorithms, CS can recover high-quality spectral images with fewer data, reducing redundancy and enhancing imaging efficiency. In 2022, Cai et al. from Tsinghua University utilised a metasurface for spectral imaging in the visible spectrum. They integrated 158,400 reconfigurable metasurface units onto a detector, where each spectral detection pixel was formed by a 5 × 5 neighbouring detector array. By using a CS spectral reconstruction algorithm, they achieved spectral imaging with a resolution of 0.8 nm [197].
In recent years, the combination of hyperspectral super-resolution algorithms with ML has further advanced the field. By learning spectral features and other prior knowledge, higher precision spectral reconstruction and resolution enhancement have been achieved simultaneously. The main focus of recent work has been on designing filters using ML methods and performing corresponding spectral reconstruction. Figure 2f illustrates a typical design framework for a parameter-constrained spectral encoder and decoder. Thanks to advancements in thin-film technology, it is now possible to reproduce the designed filter response curves, significantly improving both the speed and resolution of spectral reconstruction [198].
Up to now, in light of the advancements in detectors, dispersive components, and reconstruction algorithms, scanning imaging spectrometers operating in the 400–2500 nm range have been successfully applied in plant growth monitoring, plant health monitoring, and hyperspectral data-based breeding. In the future, snapshot spectral imaging systems with high temporal, spatial, and spectral resolution are expected to be widely adopted in smart agriculture.
Reducing Costs and Optimising Data Analysis Programs
Hyperspectral technology, bolstered by advancements in ML, has unlocked new possibilities in plant research, facilitating the creation of open HSI databases, sophisticated algorithms and accessible software packages. These innovations have streamlined data analysis, bolstering our capacity to assess and understand plant characteristics and their reactions to environmental factors. Despite its promise, the broader use of hyperspectral technology in agriculture is hindered by challenges including the complexity of data analysis, the high cost of hyperspectral cameras, scarcity of public datasets, and absence of standardised systems for specific applications [199]. To democratise access to this technology, it is imperative to develop affordable, targeted instruments and data processing solutions.
Looking ahead, the convergence of artificial intelligence and data mining offers a compelling path to system optimisation, enhancing efficiency and affordability. There is a need to concentrate on perfecting data selection protocols and advancing sensor technology to ensure that cost reduction does not affect the integrity of data. Greater insight into the technical variability present within datasets and the establishment of corrective methods for identified errors is essential. The implementation of internal standards could standardise data, fostering better integration of research findings (see Outstanding questions). The progression of HSI in plant phenotyping will rely on its adaptability and synergy with emerging technologies, offering valuable insights that propel the evolution of precision agriculture and plant breeding. This future demands our engagement, calling for creative approaches and collaborative efforts to optimally exploit the capabilities of this transformative tool.
Conclusion
This review systematically summarises hyperspectral data mining methodologies and catalogues accessible software, toolkits, and code repositories, providing a critical resource for streamlining hyperspectral data analysis. We further consolidate publicly available hyperspectral databases, addressing a key gap in plant phenomic research. While DL advances have accelerated multi-omics integration and refined trait characterisation, challenges remain in standardising data acquisition and processing protocols across scales. Our work underscores the urgent need for collaborative efforts to establish large-scale, open hyperspectral databases alongside cost-effective sensors and open-source analytical pipelines. These contributions lay a foundation for scalable, precision-driven innovations in plant breeding and agricultural sustainability.
Outstanding Questions
How can we reduce costs while maintaining HSI quality through the use of hardware solutions such as the CS-HSI, and establish standardised data collection and annotation methods to ensure consistency and comparability across datasets?
How can we optimise the acquisition process to achieve high temporal, spatial, and spectral resolution in HSI, thereby accelerating data collection to accommodate the requirements of high-throughput developmental needs in precision agriculture?
How can we further optimise and harmonise the integration of HSI with other technologies to enhance the throughput and resolution of multi-scale, dynamic biological processes, while minimising the computational and analytical challenges associated with the massive datasets generated by these high-throughput systems?
How to develop an HSI system for real-time data processing and analysis and integrate it into an agricultural decision support system to provide real-time decision basis for precision agriculture, while also lowering the threshold for hyperspectral data analysis?
Author Contributions
Jingyan Song: conceptualization (equal), writing – original draft (lead), data curation (equal), writing – review and editing (lead), visualization (lead). Haifeng Liang: conceptualization (equal), writing – original draft (lead), writing – review and editing (lead), visualization (lead). Bingjie Lu: writing – original draft (equal), data curation (equal), visualization (equal). Jing Guo: writing – original draft (equal), data curation (equal), visualization (equal). Yuan Gao: writing – original draft (equal). Xiao Hu: writing – original draft (equal). Manlin Yang: data curation (equal). Xiaofan Li: data curation (equal). Zhenyu Wang: writing – review and editing (equal). Yongqi Chen: data curation (equal). Yinyin Zhang: data curation (equal). Shen Su: data curation (equal). Zhangyun Gao: data curation (equal). Shijie Li: writing – original draft (equal), writing – review and editing (equal). Ping Chen: writing – original draft (equal). Jing Wang: supervision (lead), writing – review and editing (lead). Wanneng Yang: conceptualization (lead), funding acquisition (lead), supervision (lead), writing – review and editing (lead). Hui Feng: conceptualization (lead), funding acquisition (lead), supervision (lead), writing – review and editing (lead).
Acknowledgements
We gratefully acknowledge the support from the following funding sources: (1) The Fundamental Research Funds for the Central University (2662024SZ001); (2) The STI2030-Major Projects, Biological Breeding-National Science and Technology Major Project (2022ZD0401801); (3) National key research and development program (2022YFD1900701-4); (4) General Program of National Natural Science Foundation of China (32470432); (5) Hubei Provincial Department of Education Science and Technology Plant Project (2023DJC153); (6) Wuhan Science and Technology Plan Project (2023020402010780).
Conflicts of Interest
The authors declare no conflicts of interest.
W. Yang, H. Feng, X. Zhang, et al., “Crop Phenomics and High‐Throughput Phenotyping: Past Decades, Current Challenges, and Future Perspectives,” Molecular Plant 13, no. 2 (2020): 187–214, https://doi.org/10.1016/j.molp.2020.01.008.
P. Song, J. Wang, X. Guo, W. Yang, and C. Zhao, “High‐Throughput Phenotyping: Breaking Through the Bottleneck in Future Crop Breeding,” Crop Journal 9, no. 3 (2021): 633–645, https://doi.org/10.1016/j.cj.2021.03.015.
C. Wang, B. Liu, L. Liu, et al., “A Review of Deep Learning Used in the Hyperspectral Image Analysis for Agriculture,” Artificial Intelligence Review 54, no. 7 (2021): 5205–5253, https://doi.org/10.1007/s10462‐021‐10018‐y.
G. A. Blackburn, “Hyperspectral Remote Sensing of Plant Pigments,” Journal of Experimental Botany 58, no. 4 (2007): 855–867, https://doi.org/10.1093/jxb/erl123.
G. Y. Wu, M. A. Al‐qaness, D. Al‐Alimi, A. Dahou, M. Abd Elaziz, and A. A. Ewees, “Hyperspectral Image Classification Using Graph Convolutional Network: A Comprehensive Review,” Expert Systems with Applications 257 (2024): 125106, https://doi.org/10.1016/j.eswa.2024.125106.
L. Bian, Z. Wang, Y. Zhang, et al., “A Broadband Hyperspectral Image Sensor With High Spatio‐Temporal Resolution,” Nature 635, no. 8037 (2024): 73–81, https://doi.org/10.1038/s41586‐024‐08109‐1.
M. Faraji‐Dana, E. Arbabi, A. Arbabi, S. M. Kamali, H. Kwon, and A. Faraon, “Compact Folded Metasurface Spectrometer,” Nature Communications 9, no. 1 (2018): 4196, https://doi.org/10.1038/s41467‐018‐06495‐5.
Z. Y. Yang, T. Albrow‐Owen, W. Cai, and T. Hasan, “Miniaturization of Optical Spectrometers,” Science 371, no. 6528 (2021): 480–+, https://doi.org/10.1126/science.abe0722.
S. Wu, C. Huang, L. Yu, H. Xue, and J. Lin, “Analytical Design of an Advanced Littrow‐Offner Spectrometer Simultaneously Off‐Axis in Meridian and Sagittal Planes for Ultraviolet Imaging With High Performance,” Applied Optics 60, no. 32 (2021): 10094–10100, https://doi.org/10.1364/ao.441758.
M. Yako, Y. Yamaoka, T. Kiyohara, et al., “Video‐Rate Hyperspectral Camera Based on a CMOS‐Compatible Random Array of Fabry‐Perot Filters,” Nature Photonics 17, no. 3 (2023): 218–+, https://doi.org/10.1038/s41566‐022‐01141‐5.
M. J. Grotevent, S. Yakunin, D. Bachmann, et al., “Integrated Photodetectors for Compact Fourier‐Transform Waveguide Spectrometers,” Nature Photonics 17, no. 1 (2023): 59–64, https://doi.org/10.1038/s41566‐022‐01088‐7.
Z. Y. Yang, T. Albrow‐Owen, H. Cui, et al., “Single‐Nanowire Spectrometers,” Science 365, no. 6457 (2019): 1017–+, https://doi.org/10.1126/science.aax8814.
J. Park, X. Feng, R. Liang, and L. Gao, “Snapshot Multidimensional Photography Through Active Optical Mapping,” Nature Communications 11, no. 1 (2020): 5602, https://doi.org/10.1038/s41467‐020‐19418‐0.
Y. Su, D. Yao, S. Wang, X. Zhan, and H. Xu, “Broadband N+−N Photodiode Based on Tellurium‐Doped Black Silicon Fabricated by Femtosecond Laser Filament,” Optics & Laser Technology 123 (2020): 105913, https://doi.org/10.1016/j.optlastec.2019.105913.
Y. Gao, H. Cansizoglu, K. G. Polat, et al., “Photon‐Trapping Microstructures Enable High‐Speed High‐Efficiency Silicon Photodiodes,” Nature Photonics 11, no. 5 (2017): 301–308, https://doi.org/10.1038/nphoton.2017.37.
E. Talamas Simola, V. Kiyek, A. Ballabio, et al., “CMOS‐Compatible Bias‐Tunable Dual‐Band Detector Based on GeSn/Ge/Si Coupled Photodiodes,” ACS Photonics 8, no. 7 (2021): 2166–2173, https://doi.org/10.1021/acsphotonics.1c00617.
D. Ahn, S. Jeon, H. Suh, et al., “High‐Responsivity InAs Quantum Well Photo‐FET Integrated on Si Substrates for Extended‐Range Short‐Wave Infrared Photodetector Applications,” Photonics Research 11, no. 8 (2023): 1465–1473, https://doi.org/10.1364/PRJ.491498.
T. Dougakiuchi, K. Fujita, T. Hirohata, A. Ito, M. Hitaka, and T. Edamura, “High Photoresponse in Room Temperature Quantum Cascade Detector Based on Coupled Quantum Well Design (Vol 109, 261107, 2016),” Applied Physics Letters 110, no. 1 (2017): 109902, https://doi.org/10.1063/1.4978208.
X. Wang, M. Wang, Y. Liao, et al., “Molecular‐Beam Epitaxy‐Grown HgCdTe Infrared Detector: Material Physics, Structure Design, and Device Fabrication,” Science China Physics, Mechanics & Astronomy 66, no. 3 (2023): 237302, https://doi.org/10.1007/s11433‐022‐2003‐2.
B. A. Taha, A. J. Addie, A. C. Kadhim, et al., “Plasmonic‐Enabled Nanostructures for Designing the Next Generation of Silicon Photodetectors: Trends, Engineering and Opportunities,” Surfaces and Interfaces 48 (2024): 104334, https://doi.org/10.1016/j.surfin.2024.104334.
L. Xu, S. Shi, W. Gong, et al., “Mapping 3D Plant Chlorophyll Distribution From Hyperspectral LiDAR by a Leaf‐Canopyradiative Transfer Model,” International Journal of Applied Earth Observation and Geoinformation 127 (2024): 103649, https://doi.org/10.1016/j.jag.2024.103649.
X. Yang, R. Yang, Y. Ye, Z. Yuan, D. Wang, and K. Hua, “Winter Wheat SPAD Estimation From UAV Hyperspectral Data Using Cluster‐Regression Methods,” International Journal of Applied Earth Observation and Geoinformation 105 (2021): 102618, https://doi.org/10.1016/j.jag.2021.102618.
J. Guo, J. Zhang, S. Xiong, et al., “Hyperspectral Assessment of Leaf Nitrogen Accumulation for Winter Wheat Using Different Regression Modeling,” Precision Agriculture 22, no. 5 (2021): 1634–1658, https://doi.org/10.1007/s11119‐021‐09804‐z.
J. Wang, Y. Chu, G. Chen, et al., “Characterization and Identification of NPK Stress in Rice Using Terrestrial Hyperspectral Images,” Plant Phenomics 6 (2024): 0197, https://doi.org/10.34133/plantphenomics.0197.
R. H. Furlanetto, L. G. T. Crusiol, J. V. F. Gonçalves, et al., “Machine Learning as a Tool to Predict Potassium Concentration in Soybean Leaf Using Hyperspectral Data,” Precision Agriculture 24, no. 6 (2023): 2264–2292, https://doi.org/10.1007/s11119‐023‐10040‐w.
J. Yue, J. Wang, Z. Zhang, et al., “Estimating Crop Leaf Area Index and Chlorophyll Content Using a Deep Learning‐Based Hyperspectral Analysis Method,” Computers and Electronics in Agriculture 227 (2024): 109653, https://doi.org/10.1016/j.compag.2024.109653.
G. Tagliabue, M. Boschetti, G. Bramati, et al., “Hybrid Retrieval of Crop Traits From Multi‐Temporal PRISMA Hyperspectral Imagery,” ISPRS Journal of Photogrammetry and Remote Sensing 187 (2022): 362–377, https://doi.org/10.1016/j.isprsjprs.2022.03.014.
T. Xu, F. Wang, Z. Shi, L. Xie, and X. Yao, “Dynamic Estimation of Rice Aboveground Biomass Based on Spectral and Spatial Information Extracted From Hyperspectral Remote Sensing Images at Different Combinations of Growth Stages,” ISPRS Journal of Photogrammetry and Remote Sensing 202 (2023): 169–183, https://doi.org/10.1016/j.isprsjprs.2023.05.021.
H. Feng, Y. Fan, J. Yue, et al., “Estimation of Potato Above‐Ground Biomass Based on the VGC‐AGB Model and Deep Learning,” Computers and Electronics in Agriculture 232 (2025): 110122, https://doi.org/10.1016/j.compag.2025.110122.
B. Li, X. Xu, L. Zhang, et al., “Above‐Ground Biomass Estimation and Yield Prediction in Potato by Using UAV‐Based RGB and Hyperspectral Imaging,” ISPRS Journal of Photogrammetry and Remote Sensing 162 (2020): 161–172, https://doi.org/10.1016/j.isprsjprs.2020.02.013.
A. Kayad, F. A. Rodrigues Jr., S. Naranjo, et al., “Radiative Transfer Model Inversion Using High‐Resolution Hyperspectral Airborne Imagery—Retrieving Maize LAI to Access Biomass and Grain Yield,” Field Crops Research 282 (2022): 108449, https://doi.org/10.1016/j.fcr.2022.108449.
Y. Bai and X. Jin, “Hyperspectral Approaches for Rapid and Spatial Plant Disease Monitoring,” Trends in Plant Science 29, no. 6 (2024): 711–712, https://doi.org/10.1016/j.tplants.2024.03.011.
S. Feng, D. Zhao, Q. Guan, et al., “A Deep Convolutional Neural Network‐Based Wavelength Selection Method for Spectral Characteristics of Rice Blast Disease,” Computers and Electronics in Agriculture 199 (2022): 107199, https://doi.org/10.1016/j.compag.2022.107199.
G. Xuan, Q. Li, Y. Shao, and Y. Shi, “Early Diagnosis and Pathogenesis Monitoring of Wheat Powdery Mildew Caused by Blumeria graminis Using Hyperspectral Imaging,” Computers and Electronics in Agriculture 197 (2022): 106921, https://doi.org/10.1016/j.compag.2022.106921.
K. Wei, Y. Fu, and H. Huang, “3‐D Quasi‐Recurrent Neural Network for Hyperspectral Image Denoising,” IEEE Transactions on Neural Networks and Learning Systems 32, no. 1 (2021): 363–375, https://doi.org/10.1109/TNNLS.2020.2978756.
X. Wang, W. Yang, Y. Yang, M. Huang, and Q. Zhu, “Identification of Tomato Bacterial Wilt Severity Based on Hyperspectral Imaging Technology and Spectrum Transformer Network,” Ecological Informatics 78 (2023): 102353, https://doi.org/10.1016/j.ecoinf.2023.102353.
K. Berger, M. Machwitz, M. Kycko, et al., “Multi‐Sensor Spectral Synergies for Crop Stress Detection and Monitoring in the Optical Domain: A Review,” Remote Sensing of Environment 280 (2022): 113198, https://doi.org/10.1016/j.rse.2022.113198.
S. Paulus and A.‐K. Mahlein, “Technical Workflows for Hyperspectral Plant Image Assessment and Processing on the Greenhouse and Laboratory Scale,” GigaScience 9, no. 8 (2020): giaa090, https://doi.org/10.1093/gigascience/giaa090.
H. Chu, C. Zhang, M. Wang, et al., “Hyperspectral Imaging With Shallow Convolutional Neural Networks (SCNN) Predicts the Early Herbicide Stress in Wheat Cultivars,” Journal of Hazardous Materials 421 (2022): 126706, https://doi.org/10.1016/j.jhazmat.2021.126706.
L. Yao, M. Xu, Y. Liu, R. Niu, X. Wu, and Y. Song, “Estimating of Heavy Metal Concentration in Agricultural Soils From Hyperspectral Satellite Sensor Imagery: Considering the Sources and Migration Pathways of Pollutants,” Ecological Indicators 158 (2024): 111416, https://doi.org/10.1016/j.ecolind.2023.111416.
J. Yan and X. Wang, “Machine Learning Bridges Omics Sciences and Plant Breeding,” Trends in Plant Science 28, no. 2 (2023): 199–210, https://doi.org/10.1016/j.tplants.2022.08.018.
R. Joynson, G. Molero, B. Coombes, et al., “Uncovering Candidate Genes Involved in Photosynthetic Capacity Using Unexplored Genetic Variation in Spring Wheat,” Plant Biotechnology Journal 19, no. 8 (2021): 1537–1552, https://doi.org/10.1111/pbi.13568.
H. Zheng, W. Tang, T. Yang, et al., “Grain Protein Content Phenotyping in Rice via Hyperspectral Imaging Technology and a Genome‐Wide Association Study,” Plant Phenomics 6 (2024): 0200, https://doi.org/10.34133/plantphenomics.0200.
G. Zhang, J. Zhou, Y. Peng, et al., “High‐Throughput Phenotyping‐Based Quantitative Trait Loci Mapping Reveals the Genetic Architecture of the Salt Stress Tolerance of Brassica napus,” Plant, Cell and Environment 46, no. 2 (2023): 549–566, https://doi.org/10.1111/pce.14485.
X. Wu, H. Feng, D. Wu, et al., “Using High‐Throughput Multiple Optical Phenotyping to Decipher the Genetic Architecture of Maize Drought Tolerance,” Genome Biology 22, no. 1 (2021): 185, https://doi.org/10.1186/s13059‐021‐02377‐0.
Z. Yang, A. Li, J. Chen, et al., “Machine Learning Phenotyping and GWAS Reveal Genetic Basis of Cd Tolerance and Absorption in Jute,” Environmental Pollution 362 (2024): 124918, https://doi.org/10.1016/j.envpol.2024.124918.
A. Mangiarotti, M. Siri, N. W. Tam, Z. Zhao, L. Malacrida, and R. Dimova, “Biomolecular Condensates Modulate Membrane Lipid Packing and Hydration,” Nature Communications 14, no. 1 (2023): 6081, https://doi.org/10.1038/s41467‐023‐41709‐5.
J. P. M. Corbin, R. J. Best, I. J. Garthwaite, et al., “Hyperspectral Leaf Reflectance Detects Interactive Genetic and Environmental Effects on Tree Phenotypes, Enabling Large‐Scale Monitoring and Restoration Planning Under Climate Change,” Plant, Cell and Environment 48, no. 3 (2025): 1842–1857, https://doi.org/10.1111/pce.15263.
N. Hu, W. Li, C. Du, et al., “Predicting Micronutrients of Wheat Using Hyperspectral Imaging,” Food Chemistry 343 (2021): 128473, https://doi.org/10.1016/j.foodchem.2020.128473.
M. Medina–García, J. M. Amigo, M. A. Martínez‐Domingo, E. M. Valero, and A. M. Jiménez–Carvelo, “Strategies for Analysing Hyperspectral Imaging Data for Food Quality and Safety Issues—A Critical Review of the Last 5 Years,” Microchemical Journal 214 (2025): 113994, https://doi.org/10.1016/j.microc.2025.113994.
M. Belgiu, M. Marshall, M. Boschetti, M. Pepe, A. Stein, and A. Nelson, “PRISMA and Sentinel‐2 Spectral Response to the Nutrient Composition of Grains,” Remote Sensing of Environment 292 (2023): 113567, https://doi.org/10.1016/j.rse.2023.113567.
D. An, L. Zhang, Z. Liu, J. Liu, and Y. Wei, “Advances in Infrared Spectroscopy and Hyperspectral Imaging Combined With Artificial Intelligence for the Detection of Cereals Quality,” Critical Reviews in Food Science and Nutrition 63, no. 29 (2023): 9766–9796, https://doi.org/10.1080/10408398.2022.2066062.
Y. Wang, F. Xiong, Y. Zhang, et al., “Application of Hyperspectral Imaging Assisted With Integrated Deep Learning Approaches in Identifying Geographical Origins and Predicting Nutrient Contents of Coix Seeds,” Food Chemistry 404 (2023): 134503, https://doi.org/10.1016/j.foodchem.2022.134503.
Y. Han, Z. Liu, K. Khoshelham, and S. H. Bai, “Quality Estimation of Nuts Using Deep Learning Classification of Hyperspectral Imagery,” Computers and Electronics in Agriculture 180 (2021): 105868, https://doi.org/10.1016/j.compag.2020.105868.
N. Wu, S. Weng, J. Chen, Q. Xiao, C. Zhang, and Y. He, “Deep Convolution Neural Network With Weighted Loss to Detect Rice Seeds Vigor Based on Hyperspectral Imaging Under the Sample‐Imbalanced Condition,” Computers and Electronics in Agriculture 196 (2022): 106850, https://doi.org/10.1016/j.compag.2022.106850.
H. Qi, Z. Huang, B. Jin, et al., “SAM‐GAN: An Improved DCGAN for Rice Seed Viability Determination Using Near‐Infrared Hyperspectral Imaging,” Computers and Electronics in Agriculture 216 (2024): 108473, https://doi.org/10.1016/j.compag.2023.108473.
A. Malik, B. Ram, D. Arumugam, Z. Jin, X. Sun, and M. Xu, “Predicting Gypsum Tofu Quality From Soybean Seeds Using Hyperspectral Imaging and Machine Learning,” Food Control 160 (2024): 110357, https://doi.org/10.1016/j.foodcont.2024.110357.
J.‐Y. Song, Z. S. Qin, C. W. Xue, L. F. Bian, and C. Yang, “Fruit Grading System by Reconstructed 3D Hyperspectral Full‐Surface Images,” Postharvest Biology and Technology 212 (2024): 112898, https://doi.org/10.1016/j.postharvbio.2024.112898.
Z. Wei, H. Liu, J. Xu, Y. Li, J. Hu, and S. Tian, “Quality Grading Method for Pleurotus eryngii During Postharvest Storage Based on Hyperspectral Imaging and Multiple Quality Indicators,” Food Control 166 (2024): 110763, https://doi.org/10.1016/j.foodcont.2024.110763.
S. Wang, X. Huang, W. Han, X. Zhang, and J. Li, “Geological Remote Sensing Interpretation via a Local‐to‐Global Sensitive Feature Fusion Network,” International Journal of Applied Earth Observation and Geoinformation 135 (2024): 104258, https://doi.org/10.1016/j.jag.2024.104258.
B. Goyal, A. Dogra, S. Agrawal, B. Sohi, and A. Sharma, “Image Denoising Review: From Classical to State‐of‐the‐Art Approaches,” Information Fusion 55 (2020): 220–244, https://doi.org/10.1016/j.inffus.2019.09.003.
P. Xu, B. Chen, L. Xue, J. Zhang, L. Zhu, and H. Duan, “A New MNF–BM4D Denoising Algorithm Based on Guided Filtering for Hyperspectral Images,” ISA Transactions 92 (2019): 315–324, https://doi.org/10.1016/j.isatra.2019.02.018.
A. Asokan and J. Anitha, “Adaptive Cuckoo Search Based Optimal Bilateral Filtering for Denoising of Satellite Images,” ISA Transactions 100 (2020): 308–321, https://doi.org/10.1016/j.isatra.2019.11.008.
Q. Zhang, Q. Yuan, J. Li, F. Sun, and L. Zhang, “Deep Spatio‐Spectral Bayesian Posterior for Hyperspectral Image Non‐i.i.d. Noise Removal,” ISPRS Journal of Photogrammetry and Remote Sensing 164 (2020): 125–137, https://doi.org/10.1016/j.isprsjprs.2020.04.010.
D. Gao and X. Wu, “Multispectral Image Restoration via Inter‐ and Intra‐block Sparse Estimation Based on Physically‐Induced Joint Spatiospectral Structures,” IEEE Transactions on Image Processing 27, no. 8 (2018): 4038–4051, https://doi.org/10.1109/tip.2018.2828341.
X. Bai, F. Xu, L. Zhou, Y. Xing, L. Bai, and J. Zhou, “Nonlocal Similarity Based Nonnegative Tucker Decomposition for Hyperspectral Image Denoising,” IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing 11, no. 3 (2018): 701–712, https://doi.org/10.1109/jstars.2018.2791718.
R. Zhuge, J. Wang, Z. Xu, and Y. Xu, “Single Image Denoising With a Feature‐Enhanced Network,” Neural Networks 168 (2023): 313–325, https://doi.org/10.1016/j.neunet.2023.08.056.
Y. Chen, H. Zhang, Y. Wang, Y. Yang, and J. Wu, “Flex‐DLD: Deep Low‐Rank Decomposition Model With Flexible Priors for Hyperspectral Image Denoising and Restoration,” IEEE Transactions on Image Processing 33 (2024): 1211–1226, https://doi.org/10.1109/TIP.2024.3360902.
Q. Zhang, Q. Yuan, J. Li, X. Liu, H. Shen, and L. Zhang, “Hybrid Noise Removal in Hyperspectral Imagery With a Spatial–Spectral Gradient Network,” IEEE Transactions on Geoscience and Remote Sensing 57, no. 10 (2019): 7317–7329, https://doi.org/10.1109/tgrs.2019.2912909.
Q. Wang, L. Pang, X. Cao, Z. Tian, and D. Meng, “HIDFlowNet: A Flow‐Based Deep Network for Hyperspectral Image Denoising,” Information Fusion 125 (2026): 103459, https://doi.org/10.1016/j.inffus.2025.103459.
Q. Zhang, Y. Zheng, Q. Yuan, M. Song, H. Yu, and Y. Xiao, “Hyperspectral Image Denoising: From Model‐Driven, Data‐Driven, to Model‐Data‐Driven,” IEEE Transactions on Neural Networks and Learning Systems 35, no. 10 (2024): 13143–13163, https://doi.org/10.1109/TNNLS.2023.3278866.
J. Cao, X. Rui, L. Pang, D. Meng, and X. Cao, “LatentHSI: Restore Hyperspectral Images in a Latent Space,” Information Fusion 117 (2025): 102848, https://doi.org/10.1016/j.inffus.2024.102848.
Y.‐S. Luo, X. L. Zhao, T. X. Jiang, Y. B. Zheng, and Y. Chang, “Hyperspectral Mixed Noise Removal via Spatial‐Spectral Constrained Unsupervised Deep Image Prior,” Ieee Journal of Selected Topics in Applied Earth Observations and Remote Sensing 14 (2021): 9435–9449, https://doi.org/10.1109/jstars.2021.3111404.
X. Zhang, W. Zhang, S. Guo, et al., “UnTDIP: Unsupervised Neural Network for DEM Super‐Resolution Integrating Terrain Knowledge and Deep Prior,” International Journal of Applied Earth Observation and Geoinformation 122 (2023): 103430, https://doi.org/10.1016/j.jag.2023.103430.
Z. Gündoğar, B. U. Toreyin, and M. Demiralp, “Tridiagonal Folmat Enhanced Multivariance Products Representation Based Hyperspectral Data Compression,” Ieee Journal of Selected Topics in Applied Earth Observations and Remote Sensing 11, no. 9 (2018): 3272–3278, https://doi.org/10.1109/JSTARS.2018.2851368.
B. Carpentieri, A. Castiglione, A. De Santis, F. Palmieri, and R. Pizzolante, “One‐Pass Lossless Data Hiding and Compression of Remote Sensing Data,” Future Generation Computer Systems 90 (2019): 222–239, https://doi.org/10.1016/j.future.2018.07.051.
E. Pavez, B. Girault, A. Ortega, and P. A. Chou, “Two Channel Filter Banks on Arbitrary Graphs With Positive Semi Definite Variation Operators,” IEEE Transactions on Signal Processing 71 (2023): 917–932, https://doi.org/10.1109/TSP.2023.3257983.
L. Deng, J. Sun, Y. Chen, et al., “M2H‐Net: A Reconstruction Method for Hyperspectral Remotely Sensed Imagery,” ISPRS Journal of Photogrammetry and Remote Sensing 173 (2021): 323–348, https://doi.org/10.1016/j.isprsjprs.2021.01.019.
C. Gomes, I. Wittmann, D. Robert, et al., “Lossy Neural Compression for Geospatial Analytics: A Review,” IEEE Geoscience and Remote Sensing Magazine (2025): 2–40, https://doi.org/10.1109/MGRS.2025.3546527.
L. Xu, J. Raitoharju, A. Iosifidis, and M. Gabbouj, “Saliency‐Based Multilabel Linear Discriminant Analysis,” IEEE Transactions on Cybernetics 52, no. 10 (2022): 10200–10213, https://doi.org/10.1109/tcyb.2021.3069338.
X. Li, Q. Wang, F. Nie, and M. Chen, “Locality Adaptive Discriminant Analysis Framework,” IEEE Transactions on Cybernetics 52, no. 8 (2022): 7291–7302, https://doi.org/10.1109/tcyb.2021.3049684.
H. Nie, Q. Wu, H. Zhao, W. Ding, and M. Deveci, “Flexible Adaptive Graph Embedding for Semi‐Supervised Dimension Reduction,” Information Fusion 99 (2023): 101872, https://doi.org/10.1016/j.inffus.2023.101872.
P. Xu, Z. Deng, J. Wang, Q. Zhang, K. S. Choi, and S. Wang, “Transfer Representation Learning With TSK Fuzzy System,” IEEE Transactions on Fuzzy Systems 29, no. 3 (2021): 649–663, https://doi.org/10.1109/tfuzz.2019.2958299.
Y. Chen and X. Yang, “Online Adaptive Kernel Learning With Random Features for Large‐Scale Nonlinear Classification,” Pattern Recognition 131 (2022): 108862, https://doi.org/10.1016/j.patcog.2022.108862.
P. W. Chuang Sun, R. Yan, R. X. Gao, X. Chen, and X. Chen, “Machine Health Monitoring Based on Locally Linear Embedding With Kernel Sparse Representation for Neighborhood Optimization,” Mechanical Systems and Signal Processing 114 (2019): 25–34, https://doi.org/10.1016/j.ymssp.2018.04.044.
Yi Liu, K. Liu, J. Yang, and Y. Yuan, “Spatial‐Neighborhood Manifold Learning for Nondestructive Testing of Defects in Polymer Composites,” IEEE Transactions on Industrial Informatics 16, no. 7 (2019): 4639–4649, https://doi.org/10.1109/TII.2019.2949358.
C. Yang, S. Ma, and Q. Han, “Robust Discriminant Latent Variable Manifold Learning for Rotating Machinery Fault Diagnosis,” Engineering Applications of Artificial Intelligence 126 (2023): 106996, https://doi.org/10.1016/j.engappai.2023.106996.
J. F. Ruisheng Ran, Li Zheng, J. Wang, and B. Fang, “Locality Preserving Projections With Autoencoder,” Expert Systems with Applications 242 (2023), https://doi.org/10.1016/j.eswa.2023.122750.
H. Naeem and A. A. Bin‐Salem, “A CNN‐LSTM Network With Multi‐Level Feature Extraction‐Based Approach for Automated Detection of Coronavirus From CT Scan and X‐Ray Images,” Applied Soft Computing 113 (2021): 107918, https://doi.org/10.1016/j.asoc.2021.107918.
M. Jamali, P. Davidsson, R. Khoshkangini, M. G. Ljungqvist, and R. C. Mihailescu, “Context in Object Detection: A Systematic Literature Review,” Artificial Intelligence Review 58, no. 6 (2025): 175, https://doi.org/10.1007/s10462‐025‐11186‐x.
W. Yu, H. Huang, and G. Shen, “Deep Spectral–Spatial Feature Fusion‐Based Multiscale Adaptable Attention Network for Hyperspectral Feature Extraction,” IEEE Transactions on Instrumentation and Measurement 72 (2023): 1–13, https://doi.org/10.1109/TIM.2022.3222480.
J. Xu, K. Li, Z. Li, et al., “Fuzzy Graph Convolutional Network for Hyperspectral Image Classification,” Engineering Applications of Artificial Intelligence 127 (2024): 107280, https://doi.org/10.1016/j.engappai.2023.107280.
M. Zhang, M. Gong, H. He, and S. Zhu, “Symmetric all Convolutional Neural‐Network‐Based Unsupervised Feature Extraction for Hyperspectral Images Classification,” IEEE Transactions on Cybernetics 52, no. 5 (2022): 2981–2993, https://doi.org/10.1109/tcyb.2020.3020540.
G. Mustafa, H. Zheng, I. H. Khan, et al., “Enhancing Fusarium Head Blight Detection in Wheat Crops Using Hyperspectral Indices and Machine Learning Classifiers,” Computers and Electronics in Agriculture 218 (2024): 108663, https://doi.org/10.1016/j.compag.2024.108663.
F. Zhao, J. Zhang, Z. Meng, H. Liu, Z. Chang, and J. Fan, “Multiple Vision Architectures‐Based Hybrid Network for Hyperspectral Image Classification,” Expert Systems with Applications 234 (2023): 121032, https://doi.org/10.1016/j.eswa.2023.121032.
M. Zhang, W. Li, and Q. Du, “Diverse Region‐Based CNN for Hyperspectral Image Classification,” IEEE Transactions on Image Processing 27, no. 6 (2018): 2623–2634, https://doi.org/10.1109/tip.2018.2809606.
X. Wang, K. Tan, P. Du, B. Han, and J. Ding, “A Capsule‐Vectored Neural Network for Hyperspectral Image Classification,” Knowledge‐Based Systems 268 (2023): 110482, https://doi.org/10.1016/j.knosys.2023.110482.
Y. Ding, Z. Zhang, X. Zhao, et al., “Multi‐Scale Receptive Fields: Graph Attention Neural Network for Hyperspectral Image Classification,” Expert Systems with Applications 223 (2023): 119858, https://doi.org/10.1016/j.eswa.2023.119858.
K. C. Flynn, G. Baath, T. O. Lee, P. Gowda, and B. Northup, “Hyperspectral Reflectance and Machine Learning to Monitor Legume Biomass and Nitrogen Accumulation,” Computers and Electronics in Agriculture 211 (2023): 107991, https://doi.org/10.1016/j.compag.2023.107991.
Y. Guo, Y. Xiao, F. Hao, et al., “Comparison of Different Machine Learning Algorithms for Predicting Maize Grain Yield Using UAV‐Based Hyperspectral Images,” International Journal of Applied Earth Observation and Geoinformation 124 (2023): 103528, https://doi.org/10.1016/j.jag.2023.103528.
C. Zhang, W. Wu, L. Zhou, H. Cheng, X. Ye, and Y. He, “Developing Deep Learning Based Regression Approaches for Determination of Chemical Compositions in Dry Black Goji Berries (Lycium ruthenicum Murr.) Using Near‐Infrared Hyperspectral Imaging,” Food Chemistry 319 (2020): 126536, https://doi.org/10.1016/j.foodchem.2020.126536.
T. U. Rehman, D. Ma, L. Wang, L. Zhang, and J. Jin, “Predictive Spectral Analysis Using an End‐to‐End Deep Model From Hyperspectral Images for High‐Throughput Plant Phenotyping,” Computers and Electronics in Agriculture 177 (2020): 105713, https://doi.org/10.1016/j.compag.2020.105713.
X. Meng, Y. Bao, Y. Wang, X. Zhang, and H. Liu, “An Advanced Soil Organic Carbon Content Prediction Model via Fused Temporal‐Spatial‐Spectral (TSS) Information Based on Machine Learning and Deep Learning Algorithms,” Remote Sensing of Environment 280 (2022): 113166, https://doi.org/10.1016/j.rse.2022.113166.
R. A. Borsoi, T. Imbiriba, J. C. M. Bermudez, et al., “Spectral Variability in Hyperspectral Data Unmixing: A Comprehensive Review,” IEEE Geoscience and Remote Sensing Magazine 9, no. 4 (2021): 223–270, https://doi.org/10.1109/MGRS.2021.3071158.
Y. Liu, H. Feng, J. Yue, et al., “Improving Potato AGB Estimation to Mitigate Phenological Stage Impacts Through Depth Features From Hyperspectral Data,” Computers and Electronics in Agriculture 219 (2024): 108808, https://doi.org/10.1016/j.compag.2024.108808.
H.‐J. He, Y. Chen, G. Li, Y. Wang, X. Ou, and J. Guo, “Hyperspectral Imaging Combined With Chemometrics for Rapid Detection of Talcum Powder Adulterated in Wheat Flour,” Food Control 144 (2023): 109378, https://doi.org/10.1016/j.foodcont.2022.109378.
X. Zhou, C. Zhao, J. Sun, Y. Cao, K. Yao, and M. Xu, “A Deep Learning Method for Predicting Lead Content in Oilseed Rape Leaves Using Fluorescence Hyperspectral Imaging,” Food Chemistry 409 (2023): 135251, https://doi.org/10.1016/j.foodchem.2022.135251.
J. Peng, W. Sun, H. C. Li, et al., “Low‐Rank and Sparse Representation for Hyperspectral Image Processing: A Review,” IEEE Geoscience and Remote Sensing Magazine 10, no. 1 (2022): 10–43, https://doi.org/10.1109/MGRS.2021.3075491.
A. Maffei, J. M. Haut, M. E. Paoletti, J. Plaza, L. Bruzzone, and A. Plaza, “A Single Model CNN for Hyperspectral Image Denoising,” IEEE Transactions on Geoscience and Remote Sensing 58, no. 4 (2020): 2516–2529, https://doi.org/10.1109/tgrs.2019.2952062.
Y. Chang, L. Yan, X. L. Zhao, H. Fang, Z. Zhang, and S. Zhong, “Weighted Low‐Rank Tensor Recovery for Hyperspectral Image Restoration,” IEEE Transactions on Cybernetics 50, no. 11 (2020): 4558–4572, https://doi.org/10.1109/TCYB.2020.2983102.
P. Ray, S. S. Reddy, and T. Banerjee, “Various Dimension Reduction Techniques for High Dimensional Data Analysis: A Review,” Artificial Intelligence Review 54, no. 5 (2021): 3473–3515, https://doi.org/10.1007/s10462‐020‐09928‐0.
L. Hu, X. Chen, G. Guo, and L. Chen, “Nonnegative Matrix Factorization With Combined Kernels for Small Data Representation,” Expert Systems with Applications 208 (2022): 118155, https://doi.org/10.1016/j.eswa.2022.118155.
Y. Fei, Y. Liu, C. Jia, Z. Li, X. Wei, and M. Chen, “A Survey of Geometric Optimization for Deep Learning: From Euclidean Space to Riemannian Manifold,” ACM Computing Surveys 57, no. 5 (2025): 1–37: Article 123, https://doi.org/10.1145/3708498.
Q. Yu, M. Zhang, L. Wang, et al., “Research on Fine‐Grained Phenotypic Analysis of Temporal Root Systems—Improved YoloV8seg Applied for Fine‐Grained Analysis of In Situ Root Temporal Phenotypes,” Advanced Science 12, no. 5 (2025): 2408144, https://doi.org/10.1002/advs.202408144.
B. Graham Ram, Y. Zhang, C. Costa, et al., “Palmer Amaranth Identification Using Hyperspectral Imaging and Machine Learning Technologies in Soybean Field,” Computers and Electronics in Agriculture 215 (2023): 108444, https://doi.org/10.1016/j.compag.2023.108444.
Z. Zhao, J. Song, M. You, et al., “Hyperspectral Anomaly Detection via Cascaded Convolutional Autoencoders With Adaptive Pixel‐Level Attention,” Expert Systems with Applications 279 (2025): 127366, https://doi.org/10.1016/j.eswa.2025.127366.
J. Li, Z. Zhang, Y. Liu, R. Song, Y. Li, and Q. Du, “SWFormer: Stochastic Windows Convolutional Transformer for Hybrid Modality Hyperspectral Classification,” IEEE Transactions on Image Processing 33 (2024): 5482–5495, https://doi.org/10.1109/TIP.2024.3465038.
Y. Huang, J. Peng, N. Chen, et al., “Cross‐Scene Wetland Mapping on Hyperspectral Remote Sensing Images Using Adversarial Domain Adaptation Network,” ISPRS Journal of Photogrammetry and Remote Sensing 203 (2023): 37–54, https://doi.org/10.1016/j.isprsjprs.2023.07.009.
S. Fei, D. Xu, Z. Chen, Y. Xiao, and Y. Ma, “MLR‐Based Feature Splitting Regression for Estimating Plant Traits Using High‐Dimensional Hyperspectral Reflectance Data,” Field Crops Research 293 (2023): 108855, https://doi.org/10.1016/j.fcr.2023.108855.
J. Deng, X. Zhang, Z. Yang, et al., “Pixel‐Level Regression for UAV Hyperspectral Images: Deep Learning‐Based Quantitative Inverse of Wheat Stripe Rust Disease Index,” Computers and Electronics in Agriculture 215 (2023): 108434, https://doi.org/10.1016/j.compag.2023.108434.
S. Yu, J. Fan, X. Lu, et al., “Deep Learning Models Based on Hyperspectral Data and Time‐Series Phenotypes for Predicting Quality Attributes in Lettuces Under Water Stress,” Computers and Electronics in Agriculture 211 (2023): 108034, https://doi.org/10.1016/j.compag.2023.108034.
C. Jin, L. Zhou, Y. Pu, C. Zhang, H. Qi, and Y. Zhao, “Application of Deep Learning for High‐Throughput Phenotyping of Seed: A Review,” Artificial Intelligence Review 58, no. 3 (2025): 76, https://doi.org/10.1007/s10462‐024‐11079‐5.
N. M. Foumani, L. Miller, C. W. Tan, G. I. Webb, G. Forestier, and M. Salehi, “Deep Learning for Time Series Classification and Extrinsic Regression: A Current Survey,” ACM Computing Surveys 56 (2024): Article 217, https://doi.org/10.1145/3649448.
R. Zhao, W. Tang, M. Liu, et al., “Spatial‐Spectral Feature Extraction for in‐Field Chlorophyll Content Estimation Using Hyperspectral Imaging,” Biosystems Engineering 246 (2024): 263–276, https://doi.org/10.1016/j.biosystemseng.2024.08.008.
X. H. Wang, J. Chen, C. Richard, and D. Brie, “Learning Spectral‐Spatial Prior via 3DDNCNN for Hyperspectral Image Deconvolution,” in ICASSP 2020 ‐ 2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) (2020), 2403–2407, https://doi.org/10.1109/ICASSP40776.2020.9054539.
Y. Song, J. Zhang, Z. Liu, et al., “Deep Learning for Hyperspectral Image Classification: A Comprehensive Review and Future Predictions,” Information Fusion 123 (2025): 103285, https://doi.org/10.1016/j.inffus.2025.103285.
S. Jozdani, D. Chen, D. Pouliot, and B. Alan Johnson, “A Review and Meta‐Analysis of Generative Adversarial Networks and Their Applications in Remote Sensing,” International Journal of Applied Earth Observation and Geoinformation 108 (2022): 102734, https://doi.org/10.1016/j.jag.2022.102734.
J. Padarian, B. Minasny, and A. McBratney, “Transfer Learning to Localise a Continental Soil vis‐NIR Calibration Model,” Geoderma 340 (2019): 279–288, https://doi.org/10.1016/j.geoderma.2019.01.009.
T. U. Rehman and J. Jin, “Deep Adversarial Domain Adaptation for Hyperspectral Calibration Model Transfer Among Plant Phenotyping Systems,” Biosystems Engineering 224 (2022): 246–258, https://doi.org/10.1016/j.biosystemseng.2022.10.016.
M. Ye, Z. Meng, and Y. Qian, “Building Cross‐Domain Mapping Chains From Multi‐CycleGAN for Hyperspectral Image Classification,” IEEE Transactions on Geoscience and Remote Sensing 62 (2024): 1–17, https://doi.org/10.1109/TGRS.2024.3431460.
P. Xu, B. Chen, J. Zhang, L. Xue, and L. Zhu, “A Region‐Based Block Compressive Sensing Algorithm for Plant Hyperspectral Images,” Computers and Electronics in Agriculture 162 (2019): 699–708, https://doi.org/10.1016/j.compag.2019.05.014.
Y. M. Cai, X. Liu, and Z. Cai, “BS‐Nets: An End‐to‐End Framework for Band Selection of Hyperspectral Image,” IEEE Transactions on Geoscience and Remote Sensing 58, no. 3 (2020): 1969–1984, https://doi.org/10.1109/tgrs.2019.2951433.
X. T. Zhang, W. Huang, Q. Wang, and X. Li, “SSR‐NET: SpatialSpectral Reconstruction Network for Hyperspectral and Multispectral Image Fusion,” IEEE Transactions on Geoscience and Remote Sensing 59, no. 7 (2021): 5953–5965, https://doi.org/10.1109/tgrs.2020.3018732.
N. Zhang, Y. Pan, H. Feng, et al., “Development of Fusarium Head Blight Classification Index Using Hyperspectral Microscopy Images of Winter Wheat Spikelets,” Biosystems Engineering 186 (2019): 83–99, https://doi.org/10.1016/j.biosystemseng.2019.06.008.
J. Hao, Y. Yan, Y. Zhang, Y. Zhang, Y. Cao, and L. Wu, “A Cross‐Scale Transfer Learning Framework: Prediction of SOD Activity From Leaf Microstructure to Macroscopic Hyperspectral Imaging,” Plant Biotechnology Journal 23, no. 4 (2025): 1091–1100, https://doi.org/10.1111/pbi.14566.
S. Karpf, M. Eibl, W. Wieser, T. Klein, and R. Huber, “A Time‐Encoded Technique for Fibre‐Based Hyperspectral Broadband Stimulated Raman Microscopy,” Nature Communications 6, no. 1 (2015): 6784, https://doi.org/10.1038/ncomms7784.
Y. Long, Q. Wang, X. Tang, X. Tian, W. Huang, and B. Zhang, “Label‐Free Detection of Maize Kernels Aging Based on Raman Hyperspcectral Imaging Techinique,” Computers and Electronics in Agriculture 200 (2022): 107229, https://doi.org/10.1016/j.compag.2022.107229.
Z. Zou, Q. Wang, Q. Wu, J. Zhen, D. Yuan, et al., “Fluorescence Hyperspectral Imaging Technology Combined With Chemometrics for Kiwifruit Quality Attribute Assessment and Non‐Destructive Judgment of Maturity,” Talanta 280 (2024): 126793, https://doi.org/10.1016/j.talanta.2024.126793.
Y. Yuan, Z. Ji, Y. Fan, et al., “Deep Learning‐Assisted Fluorescence Spectroscopy for Food Quality and Safety Analysis,” Trends in Food Science & Technology 156 (2025): 104821, https://doi.org/10.1016/j.tifs.2024.104821.
Y. Zhang, T. Liu, X. Wang, et al., “Design and Preparation of Novel Magnetic Covalent Organic Framework for the Simultaneous Preconcentration and Sensitive Determination of Six Aflatoxins in Food Samples,” Food Chemistry 471 (2025): 142702, https://doi.org/10.1016/j.foodchem.2024.142702.
S.‐W. Chun, D. J. Song, K. H. Lee, et al., “Deep Learning Algorithm Development for Early Detection of Botrytis Cinerea Infected Strawberry Fruit Using Hyperspectral Fluorescence Imaging,” Postharvest Biology and Technology 214 (2024): 112918, https://doi.org/10.1016/j.postharvbio.2024.112918.
Y. Xu, L. Lu, V. Saragadam, and K. F. Kelly, “A Compressive Hyperspectral Video Imaging System Using a Single‐Pixel Detector,” Nature Communications 15, no. 1 (2024): 1456, https://doi.org/10.1038/s41467‐024‐45856‐1.
L. Wang, J. Jin, Z. Song, et al., “LeafSpec: An Accurate and Portable Hyperspectral Corn Leaf Imager,” Computers and Electronics in Agriculture 169 (2020): 105209, https://doi.org/10.1016/j.compag.2019.105209.
J. Zhang, C. Wang, C. Yang, et al., “Evaluation of a UAV‐Mounted Consumer Grade Camera With Different Spectral Modifications and Two Handheld Spectral Sensors for Rapeseed Growth Monitoring: Performance and Influencing Factors,” Precision Agriculture 21, no. 5 (2020): 1092–1120, https://doi.org/10.1007/s11119‐020‐09710‐w.
J. Jiang, H. Ji, Y. Yan, et al., “Mining Sensitive Hyperspectral Feature to Non‐Destructively Monitor Biomass and Nitrogen Accumulation Status of Tea Plant Throughout the Whole Year,” Computers and Electronics in Agriculture 225 (2024): 109358, https://doi.org/10.1016/j.compag.2024.109358.
T. Wang, S. Wu, S. Zheng, et al., “HSIS‐SIF a High‐Performance Hyperspectral Imaging Spectrometer for Solar‐Induced Chlorophyll Fluorescence of Vegetation,” Optics and Lasers in Engineering 180 (2024): 108347, https://doi.org/10.1016/j.optlaseng.2024.108347.
S. Lantin, K. McCourt, N. Butcher, et al., “SPOT: Scanning Plant IoT Facility for High‐Throughput Plant Phenotyping,” HardwareX 15 (2023): e00468, https://doi.org/10.1016/j.ohx.2023.e00468.
Y. Mao, C. H. Betters, B. Evans, et al., “OpenHSI: A Complete Open‐Source Hyperspectral Imaging Solution for Everyone,” Remote Sensing 14, no. 9 (2022): 2244, https://doi.org/10.3390/rs14092244.
A. Islam ElManawy, D. Sun, A. Abdalla, Y. Zhu, and H. Cen, “HSI‐PP: A Flexible Open‐Source Software for Hyperspectral Imaging‐Based Plant Phenotyping,” Computers and Electronics in Agriculture 200 (2022): 107248, https://doi.org/10.1016/j.compag.2022.107248.
X. Wang, Y. Zhong, L. Zhang, and Y. Xu, “Spatial Group Sparsity Regularized Nonnegative Matrix Factorization for Hyperspectral Unmixing,” IEEE Transactions on Geoscience and Remote Sensing 55, no. 11 (2017): 6287–6304, https://doi.org/10.1109/TGRS.2017.2724944.
M. Song, F. Li, C. Yu, and C. I. Chang, “Sequential Band Fusion for Hyperspectral Anomaly Detection,” IEEE Transactions on Geoscience and Remote Sensing 60 (2022): 1–16, https://doi.org/10.1109/TGRS.2021.3069734.
Y. Zhong, X. Hu, C. Luo, X. Wang, J. Zhao, and L. Zhang, “WHU‐Hi: UAV‐Borne Hyperspectral With High Spatial Resolution (H2) Benchmark Datasets and Classifier for Precise Crop Identification Based on Deep Convolutional Neural Network With CRF,” Remote Sensing of Environment 250 (2020): 112012, https://doi.org/10.1016/j.rse.2020.112012.
D. Bohnenkamp, J. Behmann, S. Paulus, U. Steiner, and A. K. Mahlein, “A Hyperspectral Library of Foliar Diseases of Wheat,” Phytopathology® 111, no. 9 (2021): 1583–1593, https://doi.org/10.1094/phyto‐09‐19‐0335‐r.
L. Shuai, Z. Li, Z. Chen, D. Luo, and J. Mu, “A Research Review on Deep Learning Combined With Hyperspectral Imaging in Multiscale Agricultural Sensing,” Computers and Electronics in Agriculture 217 (2024): 108577, https://doi.org/10.1016/j.compag.2023.108577.
M. Ryckewaert, D. Héran, C. Feilhes, et al., “Dataset Containing Spectral Data From Hyperspectral Imaging and Sugar Content Measurements of Grapes Berries in Various Maturity Stage,” Data in Brief 46 (2023): 108822, https://doi.org/10.1016/j.dib.2022.108822.
M. Ryckewaert, D. Héran, J. P. Trani, et al., “Hyperspectral Images of Grapevine Leaves Including Healthy Leaves and Leaves With Biotic and Abiotic Symptoms,” Scientific Data 10, no. 1 (2023): 743, https://doi.org/10.1038/s41597‐023‐02642‐w.
M. A. Istiak, M. M. Syeed, M. S. Hossain, et al., “Adoption of Unmanned Aerial Vehicle (UAV) Imagery in Agricultural Management: A Systematic Literature Review,” Ecological Informatics 78 (2023): 102305, https://doi.org/10.1016/j.ecoinf.2023.102305.
I. Malounas, B. Espejo‐Garcia, K. Arvanitis, and S. Fountas, “Evaluation of a Hyperspectral Image Pipeline Toward Building a Generalisation Capable Crop Dry Matter Content Prediction Model,” Biosystems Engineering 247 (2024): 153–161, https://doi.org/10.1016/j.biosystemseng.2024.09.009.
Y. Zheng, T. Zhang, and Y. Fu, “A Large‐Scale Hyperspectral Dataset for Flower Classification,” Knowledge‐Based Systems 236 (2022): 107647, https://doi.org/10.1016/j.knosys.2021.107647.
S. Van Haute, A. Nikkhah, D. Malavi, and S. Kiani, “Prediction of Essential Oil Content in Spearmint (Mentha spicata) via Near‐Infrared Hyperspectral Imaging and Chemometrics,” Scientific Reports 13, no. 1 (2023): 4261, https://doi.org/10.1038/s41598‐023‐31517‐8.
Q. Jiang, X. Zhao, T. Zhao, et al., “A Machine‐Learning–Powered Spectral‐Dominant Multimodal Soft Wearable System for Long‐Term and Early‐Stage Diagnosis of Plant Stresses,” Science Advances 11, no. 26 (2025): eadw7279, https://doi.org/10.1126/sciadv.adw7279.
C. B. Rajesh, C. V. S. S. M. Kumar, S. S. Jha, K. Ramachandran, and R. R. Nidamanuri, “In‐Situ and Airborne Hyperspectral Data for Detecting Agricultural Activities in a Dense Forest Landscape,” Data in Brief 50 (2023): 109510, https://doi.org/10.1016/j.dib.2023.109510.
L. Wan, W. Zhou, Y. He, T. C. Wanger, and H. Cen, “Combining Transfer Learning and Hyperspectral Reflectance Analysis to Assess Leaf Nitrogen Concentration Across Different Plant Species Datasets,” Remote Sensing of Environment 269 (2022): 112826, https://doi.org/10.1016/j.rse.2021.112826.
S. J. Chang, R. Chowdhry, Y. Song, et al., “HyperPRI: A Dataset of Hyperspectral Images for Underground Plant Root Study,” Computers and Electronics in Agriculture 225 (2024): 109307, https://doi.org/10.1016/j.compag.2024.109307.
X. Zhang, B. Zhang, L. Zhang, and Y. Sun, “Hyperspectral Remote Sensing Dataset for Tea Farm,” Digital Journal of Global Change Data Repository (2017), https://doi.org/10.3974/geodb.2017.03.04.V1.
X. Wu, J. Feng, R. Shang, et al., “Multi‐Task Multi‐Objective Evolutionary Network for Hyperspectral Image Classification and Pansharpening,” Information Fusion 108 (2024): 102383, https://doi.org/10.1016/j.inffus.2024.102383.
T. Depaepe, A. I. Vélez Ramirez, F. Vandenbussche, et al., “MADI: A Multispectral Automated Dynamic Imager to Monitor Plant Health,” Plant Phenomics 7, no. 2 (2025): 100040, https://doi.org/10.1016/j.plaphe.2025.100040.
K. M. Murphy, E. Ludwig, J. Gutierrez, and M. A. Gehan, “Deep Learning in Image‐Based Plant Phenotyping,” Annual Review of Plant Biology 75, no. 1 (2024): 771–795, https://doi.org/10.1146/annurev‐arplant‐070523‐042828.
E. Asante, O. Appiah, P. Appiahene, and K. Adu, “EfficientMaize: A Lightweight Dataset for Maize Classification on Resource‐Constrained Devices,” Data in Brief 54 (2024): 110261, https://doi.org/10.1016/j.dib.2024.110261.
A. Raza, Q. U. Zaman, S. Shabala, et al., “Genomics‐Assisted Breeding for Designing Salinity‐Smart Future Crops,” Plant Biotechnology Journal 23, no. 8 (2025): 3119–3151, https://doi.org/10.1111/pbi.70104.
Q. Zhou, Z. Huang, L. Liu, et al., “High‐Throughput Spike Detection and Refined Segmentation for Wheat Fusarium Head Blight in Complex Field Environments,” Computers and Electronics in Agriculture 227 (2024): 109552, https://doi.org/10.1016/j.compag.2024.109552.
H. Feng, Z. Guo, W. Yang, et al., “An Integrated Hyperspectral Imaging and Genome‐Wide Association Analysis Platform Provides Spectral and Genetic Insights Into the Natural Variation in Rice,” Scientific Reports 7, no. 1 (2017): 4401, https://doi.org/10.1038/s41598‐017‐04668‐8.
Á. Székely, T. Szalóki, M. Jancsó, J. Pauk, and C. Lantos, “Reference Field Spectrometric Data of Albino Rice Plants,” Data in Brief 54 (2024): 110319, https://doi.org/10.1016/j.dib.2024.110319.
C. M. Montes, C. Fox, Á Sanz‐Sáez, et al., “High‐Throughput Characterization, Correlation, and Mapping of Leaf Photosynthetic and Functional Traits in the Soybean (Glycine max) Nested Association Mapping Population,” Genetics 221, no. 2 (2022): iyac065, https://doi.org/10.1093/genetics/iyac065.
M. S. Przybylska, C. Violle, D. Vile, et al., “AraDiv: A Dataset of Functional Traits and Leaf Hyperspectral Reflectance of Arabidopsis thaliana,” Scientific Data 10, no. 1 (2023): 314, https://doi.org/10.1038/s41597‐023‐02189‐w.
A. Pfordt and S. Paulus, “A Review on Detection and Differentiation of Maize Diseases and Pests by Imaging Sensors,” Journal of Plant Diseases and Protection 132, no. 1 (2024): 40, https://doi.org/10.1007/s41348‐024‐01019‐4.
Y. Ge, S. Song, S. Yu, X. Zhang, and X. Li, “Rice Seed Classification by Hyperspectral Imaging System: A Real‐World Dataset and a Credible Algorithm,” Computers and Electronics in Agriculture 219 (2024): 108776, https://doi.org/10.1016/j.compag.2024.108776.
Y. C. Tian, X. Yao, J. Yang, W. Cao, D. Hannaway, and Y. Zhu, “Assessing Newly Developed and Published Vegetation Indices for Estimating Rice Leaf Nitrogen Concentration With Ground‐ and Space‐Based Hyperspectral Reflectance,” Field Crops Research 120, no. 2 (2011): 299–310, https://doi.org/10.1016/j.fcr.2010.11.002.
S. Gao, R. Zhong, K. Yan, et al., “Evaluating the Saturation Effect of Vegetation Indices in Forests Using 3D Radiative Transfer Simulations and Satellite Observations,” Remote Sensing of Environment 295 (2023): 113665, https://doi.org/10.1016/j.rse.2023.113665.
F. Li, Y. Miao, G. Feng, et al., “Improving Estimation of Summer Maize Nitrogen Status With Red Edge‐Based Spectral Vegetation Indices,” Field Crops Research 157 (2014): 111–123, https://doi.org/10.1016/j.fcr.2013.12.018.
H. Yang, F. Li, Y. Hu, and K. Yu, “Hyperspectral Indices Optimization Algorithms for Estimating Canopy Nitrogen Concentration in Potato (Solanum tuberosum L.),” International Journal of Applied Earth Observation and Geoinformation 102 (2021): 102416, https://doi.org/10.1016/j.jag.2021.102416.
S. Peanusaha, A. Pourreza, Y. Kamiya, M. W. Fidelibus, and M. Chakraborty, “Nitrogen Retrieval in Grapevine (Vitis vinifera L.) Leaves by Hyperspectral Sensing,” Remote Sensing of Environment 302 (2024): 113966, https://doi.org/10.1016/j.rse.2023.113966.
S. Junttila, T. Hölttä, N. Saarinen, et al., “Close‐Range Hyperspectral Spectroscopy Reveals Leaf Water Content Dynamics,” Remote Sensing of Environment 277 (2022): 113071, https://doi.org/10.1016/j.rse.2022.113071.
L. Chai, H. Jiang, W. T. Crow, et al., “Estimating Corn Canopy Water Content From Normalized Difference Water Index (NDWI): An Optimized NDWI‐Based Scheme and Its Feasibility for Retrieving Corn VWC,” IEEE Transactions on Geoscience and Remote Sensing 59, no. 10 (2021): 8168–8181, https://doi.org/10.1109/TGRS.2020.3041039.
Z. Wang and Z. Sun, “Optimal Vegetation Index for Assessing Leaf Water Potential Using Reflectance Factors From the Adaxial and Abaxial Surfaces,” Computers and Electronics in Agriculture 172 (2020): 105337, https://doi.org/10.1016/j.compag.2020.105337.
S. Bhadra, V. Sagan, S. Sarkar, M. Braud, T. C. Mockler, and A. L. Eveland, “PROSAIL‐Net: A Transfer Learning‐Based Dual Stream Neural Network to Estimate Leaf Chlorophyll and Leaf Angle of Crops From UAV Hyperspectral Images,” ISPRS Journal of Photogrammetry and Remote Sensing 210 (2024): 1–24, https://doi.org/10.1016/j.isprsjprs.2024.02.020.
J. C. O. Koh, B. P. Banerjee, G. Spangenberg, and S. Kant, “Automated Hyperspectral Vegetation Index Derivation Using a Hyperparameter Optimisation Framework for High‐Throughput Plant Phenotyping,” New Phytologist 233, no. 6 (2022): 2659–2670, https://doi.org/10.1111/nph.17947.
Z. Yang, J. Tian, K. Feng, X. Gong, and J. Liu, “Application of a Hyperspectral Imaging System to Quantify Leaf‐Scale Chlorophyll, Nitrogen and Chlorophyll Fluorescence Parameters in Grapevine,” Plant Physiology and Biochemistry 166 (2021): 723–737, https://doi.org/10.1016/j.plaphy.2021.06.015.
X. Zhou, W. Huang, J. Zhang, W. Kong, R. Casa, and Y. Huang, “A Novel Combined Spectral Index for Estimating the Ratio of Carotenoid to Chlorophyll Content to Monitor Crop Physiological and Phenological Status,” International Journal of Applied Earth Observation and Geoinformation 76 (2019): 128–142, https://doi.org/10.1016/j.jag.2018.10.012.
Y. Zeng, D. Hao, A. Huete, et al., “Optical Vegetation Indices for Monitoring Terrestrial Ecosystems Globally,” Nature Reviews Earth & Environment 3, no. 7 (2022): 477–493, https://doi.org/10.1038/s43017‐022‐00298‐5.
R. Sonobe and Q. Wang, “Nondestructive Assessments of Carotenoids Content of Broadleaved Plant Species Using Hyperspectral Indices,” Computers and Electronics in Agriculture 145 (2018): 18–26, https://doi.org/10.1016/j.compag.2017.12.022.
A. R. Longmire, T. Poblete, J. Hunt, D. Chen, and P. Zarco‐Tejada, “Assessment of Crop Traits Retrieved From Airborne Hyperspectral and Thermal Remote Sensing Imagery to Predict Wheat Grain Protein Content,” ISPRS Journal of Photogrammetry and Remote Sensing 193 (2022): 284–298, https://doi.org/10.1016/j.isprsjprs.2022.09.015.
R. Hernández‐Clemente, R. M. Navarro‐Cerrillo, and P. J. Zarco‐Tejada, “Carotenoid Content Estimation in a Heterogeneous Conifer Forest Using Narrow‐Band Indices and PROSPECT+DART Simulations,” Remote Sensing of Environment 127 (2012): 298–315, https://doi.org/10.1016/j.rse.2012.09.014.
P. L. Nagler, Y. Inoue, E. Glenn, A. Russ, and C. Daughtry, “Cellulose Absorption Index (CAI) to Quantify Mixed Soil–Plant Litter Scenes,” Remote Sensing of Environment 87, no. 2–3 (2003): 310–325, https://doi.org/10.1016/j.rse.2003.06.001.
D. Sun, H. Cen, H. Weng, et al., “Using Hyperspectral Analysis as a Potential High Throughput Phenotyping Tool in GWAS for Protein Content of Rice Quality,” Plant Methods 15, no. 1 (2019): 54, https://doi.org/10.1186/s13007‐019‐0432‐x.
A. K. Mahlein, T. Rumpf, P. Welke, et al., “Development of Spectral Indices for Detecting and Identifying Plant Diseases,” Remote Sensing of Environment 128 (2013): 21–30, https://doi.org/10.1016/j.rse.2012.09.019.
X. Jian, X. Cai, K. Cui, et al., “Dynamic Brain Spectrum Acquired by a Real‐Time Ultraspectral Imaging Chip With Reconfigurable Metasurfaces,” Optica 9, no. 5 (2022): 461–468, https://doi.org/10.1364/OPTICA.440013.
W. Y. Zhang, H. Song, X. He, et al., “Deeply Learned Broadband Encoding Stochastic Hyperspectral Imaging,” Light: Science & Applications 10, no. 1 (2021): 7, https://doi.org/10.1038/s41377‐021‐00545‐2.
L. Smeesters, F. Venturini, S. Paulus, et al., “2025 Photonics for Agrifood Roadmap: Towards a Sustainable and Healthier Planet,” Journal of Physics: Photonics 7, no. 3 (2025): 032501, https://doi.org/10.1088/2515‐7647/adbea9.
© 2025. This work is published under http://creativecommons.org/licenses/by/4.0/ (the "License"). Notwithstanding the ProQuest Terms and Conditions, you may use this content in accordance with the terms of the License.