1. Introduction
Hyperspectral imagers onboard CubeSats have a variety of applications since they can retrieve the spectral characteristics of ground segments of Earth along with hundreds of bands and a wide wavelength range [1,2]. Their applications include: environmental monitoring of Earth [3,4], disaster management [5,6], mineral and resource exploration [7,8,9], and coastal and marine research [10,11,12]. Among these, HyperSCAN’s primary objective is environmental monitoring of Earth, with a focus on education, image recognition, surface analysis for land-cover change monitoring, and ecosystem dynamics. The compact size and cost-effectiveness of CubeSats, combined with the advanced capabilities of hyperspectral imagers, make them a powerful tool for a wide range of applications [13,14,15].
CubeSats projects provide hands-on experience for students involving space missions for education purposes [16,17]. To fulfill student training needs, we designed and developed a hyperspectral imaging (HSI) system aboard a 12U CubeSat mission, especially for students’ education and to test experiments for their space applications. The HyperSCAN (Hyper Spectral Camera ANalyzer) is an experimental platform to train students’ skills and validate the techniques they have developed for hyperspectral imager applications in space. The training includes their integration ability for imaging/diffraction optics design, mechanical design, firmware data processing, data verifications, and data simulations.
The HyperSCAN is an educational platform to integrate college students’ ideas and skills in optical design and data processing. Design rules for HyperSCAN target one of the scientific payloads of space missions, e.g., rocket payload and CubeSat space missions. Hence, we considered the following principles when designing our HyperSCAN: (1) a rapid prototype and (2) modular design. We adopt commercial off-the-shelf (COTS) optics for a rapid prototype [18]. The HyperSCAN using COTS optical components shortens the time needed to develop the instrument and reduces the total budget for CubeSat experimenting. For modular design, we adopted COTS optical components for future replacement in other applications. Following these principles, HyperSCAN provides an experimental platform for developing remote sensing techniques to train students’ skills.
2. Scientific Purpose
Some exciting applications may include agriculture and precise farming [19,20], and environmental monitoring. For agriculture, we can use the visual band of HyperSCAN to diagnose vegetable/fruit [21], chlorophyll [22], and soil water indexes [23], etc. Especially after natural hazards (such as typhoons, floods, and landslides), we can experiment with HyperSCAN as a prototype to evaluate the production loss or financial assistance required for farms or aquaculture after a disaster [24]. Furthermore, HyperSCAN can be utilized in a marine environment to assess the spatial distribution of microscopic organisms (e.g., Phytoplankton [25]) in their visual-band spectral features. Recently, environmental pollution is an issue to be raised for natural resources, especially for air pollution [26].
3. HyperSCAN Specifications
For the broad application of environment monitoring using our developed hyperspectral imager, we consider one significant advantage in our developed HyperSCAN, high throughput: providing high throughput with F-number 2.0 by the concave grating with a large diameter of 70 mm and a focal length of 140 mm inside the HyperSCAN. The high throughput of the HyperSCAN dispersive element not only increases the aperture diameter of the front camera lens assembly but also boosts the signal-to-noise ratio of our designed HyperSCAN.
Modular design using COTS: Using COTS components allows rapid development of complex optical systems and enables calibrations with university laboratory resources. The modular design of HyperSCAN enables easy integration of various optical components, evaluation of the optical performance, and validation of its spatial quality. For example, using a similar optical path design, we could replace the diffraction element for HyperSCAN (from visible to infrared). This will allow for quick verification at the concept level. However, we have noticed that we cannot optimize the overall weight and size to a 100% perfect design for space experiments.
Compact and lightweight: The optical system consists of the front camera lens assembly, a customized slit, a concave grating, and the imager sensor with the addition of the folding mirror. The folding mirror was used to reduce the required optical path in the limited space of CubeSat. In addition, extra efforts on weight reduction were conducted in our mechanical design without the loss of mechanical strength.
Low-cost substitution of optical components for space environment test: Commercial optical components for one test only can be replaced with space-qualified ones for future opportunities in space missions. The prototype of HyperSCAN originates from the Horiba© CP140 OEM design. For space validation purposes, low-cost commercial substitutes are used to gain experience with the available optical components in an HSI system.
Using low-cost commercial optical components for space environment tests offers several advantages. First, this approach dramatically reduces test costs and leaves a budget for follow-on R&D. Second, it enables more comprehensive design validation and optimization, thereby accelerating the R&D process. Furthermore, it minimizes the wastage of expensive components in the early stages and reduces the risk of using high-cost space components in the later stages. For universities and research institutions, this approach provides an invaluable opportunity for students to deepen their understanding of optical systems, gain hands-on experience, and prepare for future contributions to space missions.
In future space missions, the use of commercially available optical components will indirectly promote more companies to start researching and developing components that are more suitable for satellite missions and reduce the R&D costs of space missions so that they can be suitable for small businesses and schools to develop.
HyperSCAN was designed as an educational platform with a limited budget, targeted to meet the needs of academic research and education. Unlike commercial hyperspectral imaging (HSI) systems, which start at USD 28k [27] and are often prohibitively expensive, HyperSCAN aims to provide a low-cost alternative. By using affordable, commercially available optical components for testing, the development aligns with budget constraints while enabling iterative design and validation. This strategy preserves valuable resources for later development stages and offers students and researchers hands-on experience with HSI systems at a fraction of the cost. The overall development cost of HyperSCAN is approximately USD 20k, which is about 70% of the cost of comparable commercial HSI systems.
Figure 1 shows the mechanism layout of HyperSCAN and Table 1 lists the specifications of the HyperSCAN flight model. HyperSCAN in Figure 2 was integrated into a CubeSat mission named SCintillation and IONsophere eXtended (SCION-X). A mission orbit was designed at altitude of 500 km and the corresponding observation parameters are listed in Table 2.
Figure 2 shows a rendering of the structural design of the SCION-X 12U CubeSat. The green panels represent the solar panels, which are positioned opposite the nadir direction and perpendicular to the CubeSat’s flight path, indicated by a black arrow. The blue dashed-line cone illustrates the viewing direction of HyperSCAN, which is aligned parallel to the nadir direction for ground observation of Earth.
Table 3 presents a comparison of HyperSCAN with other hyperspectral imagers on CubeSats. HyperSCAN stands out with its superior spectral resolution and spectral-band number (162), enabling more detailed spectral analysis. Its size (<4 U) makes it highly suitable for CubeSats, while its low power consumption (<5 W) is significantly lower compared to imagers like HyperScout-2 (9 W) and MicroHyperspec A-series (6.6 W). Although HyperSCAN’s wavelength range (450–850 nm) is narrower than other imagers, it remains effective for applications within the visible and near-infrared spectrum. This balance of high spectral resolution, compact design, and energy efficiency positions HyperSCAN as a competitive choice for CubeSat missions requiring hyperspectral imaging.
3.1. Hardware
Figure 3 illustrates the light trajectory inside the HyperSCAN. To easily substitute the optical components in HyperSCAN, we subdivide the complex optical system into the critical optical components: the front camera lens assembly, the slit in the focal plane, the folding mirror, diffractive element, and the imaging sensor. Modular design principles increase the flexibility in versatile development and future applications, e.g., the infrared band. First, the landscape on Earth is projected through the front camera lens assembly and is focused on the slit (8 mm in length and 14.8 μm in width) at the focal plane of the camera. The reflective mirror reduces the concave grating’s required focal length (140 mm) by folding the optical path. The concave mirrors with grating focus the diffractive light into the image sensor.
Figure 4 shows the mission concept of HyperSCAN. The HyperSCAN adopts the push-broom method to scan the Earth’s surface. It concatenates a serial of sliced images along the track and composites the 3D data cube. The generated hyperspectral data cube has 2D spatial dimensions and 1D spectral dimensions. The spatial information includes the horizontal spans (80 km) and the along-track length (a total of 128 frames for data cubic).
Figure 5 shows a reconstructed RGB image by using three selected bands to reconstruct and show the single spectrum curves of different materials, including grass, sky, buildings, and glass. The image is captured by HyperSCAN and the regions of interest are highlighted.
The parts in the front camera lens assembly of HyperSCAN include (a) the external baffles, (b) the filter house, (c) the lens system, (d) the slit assembly, and (e) the mounting base. The external baffles prevent targeted light from mixing stray light, which reduces the noise from other targets. We prepare the filter house for band-pass requirements for broad applications. The lens system consists of four lenses with a focal length of f = 50 mm and f/1.8, as Figure 6 shows. The lens has a spherical symmetry, which is low-cost and easy to manufacture. The slit has a height of 8 mm and a width of 14.4 μm, equivalent to 3 pixels on the image sensor. The mounting base is mounted on the entrance of HyperSCAN. All parts are individually functioned and prepared for integration into future custom design use in other applications.
The concave grating is from an online catalog and has a focal length of 140 mm and a high throughput with f-number 1.4 [34]. The flat-field design for concave grating minimizes the aberration on the imager sensor. Table 4 show the specification of imaging sensor Python 5K (onsemi©, Phoenix, AZ, USA), which is selected from a catalog of the PYTHON series [35]. The pixel size of the CMOS sensor is 4.8 × 4.8 μm. The CMOS sensor supports correlated double sampling readout to reduce noise and increase the dynamic range. It features on-chip 10-bit A/D converters and programmable gain amplifiers, allowing for the resetting of the gain parameters and the integration time based on user needs, or the enabling of the on-chip automatic-exposure control loop to dynamically adjust these parameters without visible artifacts. The black level of images is automatically calibrated or can be fixed by adding a user-programmable offset. Considering the CubeSat data volume, we down sampled the spatial resolution and spectral resolution to the 162 bands by binning sixteen pixels in the spatial domain and decreased the spatial resolution with by a rate of 4 by binning four pixels. The original HyperSCAN spatial resolution with 48 m is downscaled to 192 m for an orbit of 500 km.
3.2. Architecture of Instrument Control and Data Handling
The main functionality of the instrument control and data handling (ICDH) board is to control a CMOS imager sensor, to process and format hyperspectral data, and to regularly record housekeeping status of the payload. Their required functions for ICDH are listed below:
Controlling the CMOS sensor via the internal links.
Receiving, formatting, and packetizing data from the CMOS sensor via the internal links.
Interfacing as a remote terminal of the onboard data handling (OBDH) subsystem via the RS422 for telecommand/telemetry.
Interfacing to the CubeSat OBDH for sending science data packets.
Connecting to the CubeSat electrical power subsystem (EPS) to receive the supply voltage.
Controlling the power supply to the CMOS sensor and other subsystems.
Controlling the heaters based on the signal received from the temperature sensors, through the service module via the internal links.
Sending a housekeeping packet for monitoring temperatures, current, voltage, and current status of the payload.
The ICDH consists of the main processing unit for the command and data handler, and the auxiliary processing unit for data compression. We selected the Microchip© (original Microsemi©, Chandler, AZ, USA) SmartFusion 2 M2S060 for the main processing unit and Texas Instruments© TMS 570 for data compression. The SmartFusion 2 M2S060 is an Arm Cortex-M3 processor with embedded flash. The SerDes devices consume as little as 13 mW/Gbps per lane, effectively reducing power consumption. It supports single-event upset immunity, ensuring data accuracy in space environments. The devices include DDR3 memory controllers that provide high-speed memory interfaces. Their relevant specifications are listed in Table 5 and Table 6 for the main processing unit and the auxiliary processing unit, respectively.
The functional blocks of the subsystem in the ICDH are shown in Figure 7. The hyperspectral data were acquired from the CMOS sensor, transferred to be compressed in the auxiliary processing unit, and stored in the external flash with a synchronous communication protocol (serial peripheral interface, SPI). The ICDH packetizes the status of voltage, current, and temperature as housekeeping data after receiving the packages from the temperature and current sensor units with I2C. The science data and housekeeping data are sent to the CubeSat on-board data handling (OBDH) via the RS422 until the ICDH receives commands from OBDH. This ICDH receives a power supply of 5 V via the CubeSat EPS interface.
Figure 8 shows the data interface between subsystems is shown in, while Figure 9 shows the power supply interface in the ICDH. The main processing unit is the control center of the ICDH. It controls the data of observation by sending commands, monitors the status of the instrument, processes science data in DDR3, transfers data to compressed in the auxiliary processing unit, and stores data in the external flash. It provides the main clock signal of the system to SPI devices such as the CMOS sensor, auxiliary processing unit, and external flash to synchronize the system of the ICDH.
The ICDH generally performs the following steps:
(1). The CMOS sensor starts to record when the OBDH sends the recording command to it through the main processing unit. One observation (OBS) takes 128 images. The main processing unit receives the digital image data generated by the visible band spectrometer after filming.
(2). The auxiliary processing unit compresses OBS data from the main processing unit using LZO compression, a lossless data compression algorithm designed for fast decompression. The LZO compression algorithm is easily implemented and compatible with the HyperSCAN system. It is thread-safe and lossless, providing faster compression speeds than the DEFLATE compression algorithm, a widely used method for text and image compression in standard formats, such as ZIP files, PNG images, and HTTP compression. In our case, the LZO algorithm reduced data transfer time by approximately 10% to 50% without any issues. Its safety and ease of implementation make LZO a preferred choice. Once compression is complete, corresponding identifiers are assigned to the OBS data and images.
(3). The external flash of the main processing unit stores the compressed data in the planned location with ECC protection. When the operation of storage is finished, the main processing unit sends an interrupt signal to the OBDH as a completed response. One is an external flash with a capacity of 64 OBS datasets and the other is a spare.
(4). The OBDH sends the reading command to the main processing unit, which packetizes the stored data as per the Consultative Committee for Space Data Systems (CCSDS) standards and returns it to the OBDH. This is a typical data acquisition procedure.
3.3. Flight Software Onboard the ICDH
We chose FreeRTOS© as our real-time-operation system inside the ICDH. FreeRTOS is an open source, real-time operating system for microcontrollers that makes small, low power edge devices, which are easy to program, deploy, secure, and manage. Each task is concurrent as the switch of timesharing between tasks runs so quickly that it might appear to run simultaneously. The tasks and their function description are listed below for the main processing unit and the auxiliary processing unit, respectively.
For the main processing unit (SF2 M2S060),
-
Command interface: to receive commands from CubeSat’s OBDH and to send the scientific data, payload health, and status packets.
-
DMA: to transfer the high-speed CMOS data from the sensor to SF2 M2S060.
-
USB: debugger in ground test phase.
-
Console: to debug and monitor the status of the ICDH.
-
Housekeeping: to regularly record the health and status of the payload.
For the auxiliary processing unit (TMS 570),
-
Data compression: to use the LZO real-time data compression library for encoding data.
-
SPI data interface: to communicate the data packet between SF2 M2S060 and TMS 570.
-
Console: to debug and monitor the message from TMS570.
4. Spatial-Resolution-Boost Experiments Using HyperSCAN Data
With aide of higher spatial resolution multispectral images (MSI), we can boost the spatial resolution of hyperspectral images (HSI) from HyperSCAN data. CNN methods are widely used to capture spatial and spectral features of hyperspectral and multispectral images, forming the backbone of spatial-resolution-boost experiments. Concepts from methods like EGST (enhanced geometric structure transformer) [36], which emphasizes multidimensional feature representation, and BBS [37], which utilizes structural constraints, provide valuable inspiration for understanding spatial-spectral relationships.
We conducted a pre-flight experiment to boost its spatial resolution. The higher spatial resolution MSI can be obtained from remote sensing data (e.g., Landsat and Sentinel mission), but due to the size constraints of CubeSats, the focal length of hyperspectral imagers is limited, which restricts the spatial resolution of data collected by CubeSats. Therefore, in this experiment, MSI data with high spatial resolution is fused with HSI data with high spectral resolution but low spatial resolution to generate HSI data with both high spatial and spectral resolution, overcoming the spatial resolution limitations of CubeSat-based hyperspectral imaging. For the CubeSat flight mission, the front camera lens assembly of HyperSCAN onboard CubeSat is designed to focus at infinity for the orbit height of 500 km. We acquired the data from the side viewing of HyperSCAN mounted on the rotator motor with varying angles. To evaluate the performance of the acquired HyperSCAN data, we selected a building on the campus of National Central University as our scanning test. We reconstructed higher spatial resolution HSI using deep learning models. The multi-sensor fusion model synthesizes the heterogeneous datasets through combining information from multiple sensors. The multi-sensor fusion model will take advantage of both the higher spatial resolution of MSI and the higher spectral resolution of HSI. In the outdoor experiment, we will adopt the best performance algorithm after comparing their evaluation metrics between scenarios of multi-sensor fusion algorithms. Furthermore, we selected five datasets including Urban, Pavia University, Pavia Centre, Botswana, and Indian Pines, which are commonly used in spatial-resolution-boost experiments, to validate the model performance. All datasets referenced in this study are available and can be downloaded from
HyperSCAN building: The HyperSCAN building dataset, acquired using the HyperSCAN system, serves as the primary dataset for validation. The data were collected at National Central University, with wavelengths ranging from 450 nm to 850 nm, and 162 valid spectral bands.
Urban: The Urban dataset, obtained by the Hydice sensor, is one of the most commonly used hyperspectral image datasets, with wavelengths ranging from 400 nm to 2500 nm. Due to dense water vapor and atmospheric effects, some bands were removed, leaving 162 bands as valid bands. The image size is 307 × 307 pixels, and the spatial resolution is 2 m.
Pavia University: The Pavia University dataset was acquired by the ROSIS sensor over Pavia University, Italy, with a range of 430 nm to 860 nm, 103 valid bands, and an image consisting of 610 × 340 pixels with a spatial resolution of 1.3 m.
Pavia Centre: The Pavia Centre dataset was acquired by ROSIS sensors, as was the Pavia University dataset. The range is of 430 nm to 860 nm, with 102 valid bands, and the image consists of 1096 × 1096 pixels with a spatial resolution of 1.3m.
Botswana: The Botswana dataset was acquired by the Hyperion sensor on NASA’s EO-1 satellite and contains images of swamps and dry woodlands. The images cover 242 bands from 400 to 2500 nm, leaving 145 valid bands after removing noisy bands and water absorption-effecting bands. The image size is 1476 × 256 and the spatial resolution is 30 m.
Indian Pines: The Indian Pines dataset was acquired by AVIRIS sensor, and the images were taken at the Indian Pines test site in north-western Indiana, covering 224 bands from 400 to 2500 nm, leaving 200 valid bands after removing noise and moisture effects. The image size is 145 × 145 and the spatial resolution is 20 m.
Figure 10 shows the two frameworks and approaches using deep learning models: (a) feature-input fusion and (b) input-level fusion. The feature-input fusion extracts spatial and spectral features independently from multiple sources (e.g., HSI and MSI), combines these features, and predicts their spatial-resolution-boost results. The input-level fusion integrates data from multiple sensors (e.g., HSI and MSI) into a single input before two sequential function blocks reconstruct them on spectral and spatial domains with a deep learning model. The feature-input fusion method better preserves and leverages the unique strengths of both hyperspectral and multispectral data. It is more adaptable to the diverse characteristics of the input data and enhances the model’s ability to handle noise and artifacts. It facilitates more precise feature integration at higher network layers, tailored to the specific task. We adopted the two-branches convolutional neural network (TBCNN) [38] for feature-input fusion and cross-mode information fusion [39] for input-level fusion. For the training data, we prepared two datasets: a lower spatial resolution using a down sampling factor of 4 for the hyperspectral images (LS-HSI) and a lower spectral resolution for the multispectral images (LS-MSI) using eight broadly selected Landsat bands. For the feature-input fusion and input-level fusion, the deep learning method can be interpreted as (a) a parallel convolution neural network (CNN) feature-extraction in HSI and MSI using a convolutional neural network [38] and (b) the sequential reconstruction of HSI and MSI [39], respectively.
For feature-input fusion, the two-branches convolutional neural network (TBCNN) implements two feature extraction methods for the spatial data and spectral data in parallel. We used data augments for prior data processing to increase the training dataset and to avoid overfitting problems. The extracted features in spatial and spectral training data were concatenated and fed to a fully connected (FC) layer, as an output of reconstructed data [38,40]. The feature layers extract characteristic values from the spatial domain in MSI and the spectral domain in HIS. The retrieved characteristic values in the spectral and spatial domains are concatenated and merged into a fully connected layer to generate the HSI-MSI data using the minimizing mean square errors. The loss function is optimized using the adaptive moment estimation using the infinity norm (ADAMAX) method with standard backward error propagation. ADAMAX is more efficient than ADAM (adaptive moment estimation) [41]. Based on the infinity norm (max-norm), ADAMAX is designed to improve the robustness and performance of the original ADAM algorithm, particularly in scenarios with large gradients.
We selected one of fusion methods as the benchmark compared to data-input fusion reconstruction, e.g., the spatial-spectral reconstruction network (SSRNET) where SSRNET is one of the efficient cross-mode information fusion methods [39]. The raw data from HSI and MSI are interpolated and interlaced for data-input fusion. Through the spatial-spectral reconstruction network, we adjusted the fusion data by making data reconstruction on the spatial and spectral domain in series. The reconstruction methods are optimized with the spatial edge loss and spectral edge loss. Otherwise, the spatial reconstruction network (SpatCNN) optimizes spatial edge loss while the spectral reconstruction network (SpecCNN) optimizes spatial edge loss [39]. The tensor fusion network (TFNet) [42] and residual tensor fusion network (ResTFNet) [42] are other data-input fusion applications. These networks often leverage advanced deep learning techniques, such as tensor fusion and residual learning, to enhance performance. The ConSSFCNN (contextual spatial-spectral feature convolutional neural network) [43] and SSFCNN (spatial-spectral feature convolutional neural network) [43] are specialized for effectively capturing and utilizing both spatial and spectral features to improve their performance. Finally, MSDCNN (multi-scale dense convolutional neural network) [44] is an architecture designed to capture multi-scale features in image processing tasks. It integrates multi-scale feature extraction with dense connectivity to improve the representation and utilization of features at different scales. Figure 11 shows reconstructed data using TBCNN, SSRNET, SpatCNN, SpecCNN, ConSSFCNN, SSFCNN, ResTFNet, TFNet, and MSDCNN where the first row indicates the RGB pseudo color (R at 650 nm, G at 530 nm, and B at 470 nm) and pixel value differences between the reconstructed LR-HSI/LR-MSI and the original HSI at 850 nm in the second row and 450 nm in the third row.
Table 7 compares the TBCNN reconstructed data with other deep learning models for evaluating data reconstruction techniques, which will be applied to future CubeSat hyperspectral missions. In Table 7, we adopted four generally used evaluating indexes, peak signal-to-noise ratio (PSNR), root mean squared error (RMSE), erreur relative globale adimensionnelle de synthèse (ERGAS) [45], and spectral angle mapper (SAM) [46]. PSNR evaluates the spatial quality of reconstructed data, reflecting the overall accuracy of pixel-intensity reconstruction, with higher values indicating better quality. RMSE measures the pixel variations between reconstructed data and simulated data, with lower values indicating greater similarity to the reference image. SAM is generally utilized to evaluate the spectral information preservation degree at each pixel [46] by calculating the spectral angular difference at each pixel, with smaller values denoting higher spectral coherence. ERGAS estimates the quality of high-resolution synthesized images by normalizing the reconstruction error to each band and accounting for spatial resolution differences, where lower values indicate higher quality [45].
In Table 7, higher PSNR, lower RMSE, lower ERGAS, and lower SAM values represent better image quality. This analysis helps understand each scheme’s strengths and weaknesses, guiding the choice of algorithms for specific applications for HyperSCAN data. The TBCNN has the highest PSNR and lowest RMSE and ERGAS, indicating superior image quality and reconstruction accuracy—the SAM value, although not the lowest, is still relatively good. SSRNET shows strong performance with the second-highest PSNR and low RMSE and ERGAS values. It has the lowest SAM value, indicating the best spectral accuracy. SpatCNN has moderate performance with average values across all metrics. It performs significantly worse than TBCNN and SSR-NET but better than SpecCNN and ConSSFCNN. SpecCNN performs poorly with low PSNR and high RMSE, ERGAS, and SAM values. This indicates poor image quality and spectral accuracy. ConSSFCNN shows the worst performance across all metrics, with the lowest PSNR and highest RMSE, ERGAS, and SAM values. ConSSFCNN is the least effective method among those compared.
The superior performance of TBCNN compared to other models can be attributed to its deeper convolutional layers, which allow for capturing detailed features. Unlike simpler architectures, such as SSRNET, which may overlook critical information due to its simpler convolution layers, TBCNN employs a more complex architecture that effectively extracts both spectral and spatial features. Additionally, TBCNN separates the processing of hyperspectral and multispectral images, enabling a more targeted extraction of spectral and spatial features. This separation ensures that each type of data are handled to maximize their unique properties, leading to improved reconstruction accuracy and overall image quality.
We selected the TBCNN as one of the candidates among various deep learning models since TBCNN has the best reconstruction performance. Figure 12 shows the concept ideas for HSI-MSI fusion data processing while Figure 13 shows the reconstructed data at selected bands (450, 530, 610, 690, 770, and 850 nm) in comparison with input data (LS-HSI) and original data (HS-HSI). The low-cost hyperspectral imager from COTS provides a unique opportunity for space applications in the future. Our designed HyperSCAN is one of the feasible ways, which is also verified in the spatial-resolution-boost experiments. Such CubeSat hyperspectral application may pave the way for future wide applications of satellite missions.
To further validate the performance of TBCNN, we conducted experiments using various datasets, including Urban, Pavia University, Pavia Centre, Botswana, and Indian Pines. Figure 14, Figure 15, Figure 16, Figure 17 and Figure 18 show the RGB pseudocolor results of TBCNN and comparison models for each dataset. Table 8, Table 9, Table 10, Table 11 and Table 12 show the evaluation indicators for TBCNN and comparison models, reflecting the quality of reconstruction images.
In comparison of the model performance, SSRNET, TFNET, MSDCNN, SSFCNN, ResTFNet, and SpatCNN show stable performance in both spectral and spatial resolution improvement, SpecCNN has an excellent performance in spatial resolution, but poor performance in spectral performance, and ConSSFCNN’s performance is less stable in both spatial and spectral. Results demonstrate that TBCNN achieves superior reconstruction performance on spectral and spatial resolution across various datasets, indicating its potential as a benchmark model for enhancing spatial resolution in HyperSCAN data in future applications.
There are still potential disadvantages to the spectrometer. The first is the use of commercial products. Although the cost and time of designing a diffraction grating is reduced, more is needed to achieve the best results in terms of size and performance. A modular design reduces design time and cost, but it loses specific observation bands, which is a trade-off that we must consider. However, it reduces the difficulty of the design and reduces the risk of failure in student experiments.
5. Conclusions
HyperSCAN is an experimental platform to validate our techniques for future hyperspectral imager applications in space. The techniques include the integration ability for imaging/diffraction optics design, mechanical skill, firmware data processing, and performance testifying. For modular design for future applications, we adopt the commercial off-the-shelf (COTS) grating component as the base to develop our hyperspectral imager. Based on required visible wavelength range, we designed and implemented the optical mechanical system, including integrating the front camera lens assembly, folding mirror, concave grating, and sensor within a compact size using optical path folding. Without data volume limits of a CubeSat, the spatial and spectral resolution of HyperSCAN can be achieved as 48 m for a front camera lens assembly with a focal length of 50 mm and a bandwidth FWHM ~10 nm (162 bands). However, due to transmission constraints from the university ground station’s limited downlink capacity, the achievable spatial resolution and data volume are restricted. To mitigate this limitation, existing high-spatial-resolution MSI data (e.g., Landsat and Sentinel missions) are integrated with the collected HSI data to enhance both spatial and spectral resolution. In addition, the spatial-resolution-boost experiments validate that the two-branches convolutional neural network (TBCNN) for feature-input fusion can greatly improve spatial resolution along with higher spatial resolution.
Conceptualization, C.-L.K.; methodology, W.-Q.C. and H.-Y.H.; software, W.-Q.C. and H.-Y.H.; validation, W.-Q.C. and H.-Y.H.; formal analysis, W.-Q.C. and H.-Y.H.; investigation, T.-H.L., L.C.C., C.-K.C. and J.-Y.L.; data curation, M.W.S. and P.-Y.L.; writing—original draft preparation, W.-Q.C., P.-Y.L. and C.-L.K.; writing—review and editing, C.-L.K. and L.C.C.; visualization, W.-Q.C. and H.-Y.H.; project administration, J.-Y.L. All authors have read and agreed to the published version of the manuscript.
The raw data supporting the conclusions of this article will be made available by the authors on request. The data are not publicly available due to privacy.
The authors would like to thank the assistance and support from the Department of Research and Development, Liscotech System Company Ltd., Taipei, Taiwan, especially from Jack Chen and Shiuan-Hal Shiu.
The authors declare no conflicts of interest.
Footnotes
Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content.
Enlarge this image.Figure 3. The HyperSCAN is composed of the following optical components: (1) the front camera lens assembly, (2) the slit in the focal plane, (3) the folding mirror, (4) diffractive element (concave grating), and (5) the imaging sensor.
Enlarge this image.Figure 4. (a) The slice scans of the Earth surface with a swath width of 80.3 km (red double-arrow line), and its along-track length of 33.3 km, which is illustrated by a red thick arrow. A black thick arrow indicates a flying direction of HyperSCAN at the CubeSat orbit height H, (b) 2D sliced image with spatial and spectral information of Earth’s surface, and (c) 3D hyperspectral data cube composed by successive line scans of ground view.
Enlarge this image.Figure 5. (a) An RGB image reconstructed from three selected bands of a hyperspectral image captured during the HyperSCAN experiment, with regions of interest highlighted with green, blue, red, and cyan filled circles. (b) The spectral radiance curves of different materials, including grass, sky, buildings, and glass, over the wavelength range of 450 nm to 850 nm.
Enlarge this image.Figure 6. The mechanical design of the camera at (a) the transparent view and (b) the 3D view. At the transparent view, a lens system of camera has 4 elements (labeled by G1, G2, G3, and G4).
Enlarge this image.Figure 10. Two frameworks and approaches using deep learning data reconstruction models: (a) feature-input fusion and (b) input-level fusion.
Enlarge this image.Figure 11. Spatial-resolution-boost experiments from low-spectral-resolution MSI and low-spatial-resolution HSI data using TBCNN, SSRNET, SpatCNN, SpecCNN, ConSSFCNN, SSFCNN, ResTFNet, TFNet, and MSDCNN where TBCNN demonstrates the minimum differences at two edge wavelength (450 and 850 nm).
Enlarge this image.Figure 12. Spatial-resolution-boost experiments using low-spectral-resolution multispectral and hyperspectral data.
Enlarge this image.Figure 13. Reconstructed data at various wavelength from TBCNN in the third row in comparison with low-spatial-resolution HSI in the first row and original high-spatial-resolution HSI in the second row. The TBCNN greatly improves the spatial resolution of low-spatial-resolution HSI along with low-spectral-resolution MSI.
Enlarge this image.Figure 14. Spatial-resolution-boost experiments on the Urban dataset: RGB pseudocolor visualization of TBCNN, SSRNET, SpatCNN, SpecCNN, ConSSFCNN, SSFCNN, ResTFNet, TFNet, and MSDCNN.
Enlarge this image.Figure 15. Spatial-resolution-boost experiments on the Pavia University dataset: RGB pseudo-color visualization of TBCNN, SSRNET, SpatCNN, SpecCNN, ConSSFCNN, SSFCNN, ResTFNet, TFNet, and MSDCNN.
Enlarge this image.Figure 16. Spatial-resolution-boost experiments on the Pavia Centre dataset: RGB pseudocolor visualization of TBCNN, SSRNET, SpatCNN, SpecCNN, ConSSFCNN, SSFCNN, ResTFNet, TFNet, and MSDCNN.
Enlarge this image.Figure 17. Spatial-resolution-boost experiments on the Botwana dataset: RGB pseudocolor visualization of TBCNN, SSRNET, SpatCNN, SpecCNN, ConSSFCNN, SSFCNN, ResTFNet, TFNet, and MSDCNN.
Enlarge this image.Figure 18. Spatial-resolution-boost experiments on the Indian Pines dataset: RGB pseudocolor visualization of TBCNN, SSRNET, SpatCNN, SpecCNN, ConSSFCNN, SSFCNN, ResTFNet, TFNet, and MSDCNN.
The specifications of HyperSCAN.
| Parameter | Value |
|---|---|
| Size | Less than 20 × 20 × 10 cm3 |
| Weight | 1.4 kG |
| Power | Less than 5 W |
| Data interface | RS422 × 2 |
| Onboard data volume | DDR w/ECC 1 1GB |
| Non-volatile data volume | Flash 2GB |
1 ECC (Error Correction Code).
The relevant parameters of HyperSCAN data for an orbit height of 500 km 1.
| Parameter | Value |
|---|---|
| Scanning mode | Push-broom |
| FOV 2 | 9.2° |
| iFOV 3 | ~4 × 10−4 rad |
| Spatial resolution | ~192 m |
| Frame rate | ~30 fps for high speed |
| Wavelength | 450–850 nm |
| Band number | 162 for binning mode |
| Data cube format | 512 × 128 pixels for 162 bands |
| Data resolution | 10 bits |
1 The orbit height varies for different applications. 2 FOV corresponds to front-camera-lens assembly with a focal length of 50 mm where the ground spatial resolution will be enhanced for using a telescope with a longer focal length. 3 The spatial resolution depends on the maximal value of the slit width (14.8 μm) and binning four pixel (4.8 μm × 4 = 19.2 μm).
Comparison of HyperSCAN with other hyperspectral imagers on CubeSats (Note *: Each image records only 3 bands, but the imager can flexibly change the bands being recorded. The symbol “-” represents the data of instrument specification which the literature does not mention).
| SCION-X | Cosine | Headwall | Aalto-1 | SAC | NASA-JPL | |
|---|---|---|---|---|---|---|
| Scanning mode | Push-broom | Push-broom | Push-broom | - | Push-broom | - |
| Wavelength [nm] | 450–850 | 400–1000 | 400–1000 | 500–900 | - | 350–1700 |
| Band number | 162 | 45 | 324 | 3 * | 150 | - |
| Size | <4 U | 1.5 U | <2 U | <1 U | 2 U | 4 U |
| Power [W] | <5 | 9 | 6.6 | - | 3.5 | - |
| Weight [kg] | 1.4 | 1.3 | 0.7 | 0.6 | 1.35 | 1.84 |
| Data resolution [bits] | 10 | 12 | 12 | - | - | 14 |
| Data cube format | 512 × 128 pixels for 162 bands | 4000 × 1850 pixels for 45 bands | 1004 pixels for 324 bands | 512 × 512 pixels for 3 * bands | - | - |
| FOV (°) | 9.2 | 23 × 16 | - | 10 | - | 10 |
The specification of the CMOS sensor 1.
| Parameter | Value |
|---|---|
| Type | Python 5K (Monochrome) |
| Optical Size | 1 inch |
| Pixel number | 2592 × 2048 |
| Pixel size | 4.8 × 4.8 μm |
| Resolution | 10 bits |
| Data output | LVDS × 4 |
| Command interface | SPI |
| Operation temperature | −40 to 85 °C |
| Power | 0.915 W |
1 The detailed information can be obtained from the website [
The parts in the main processing unit (Microchip SmartFusion 2 M2S060).
| Parts | Description |
|---|---|
| Main processing unit | Microchip SmartFusion 2 M2S060 |
| External RAM | DDR3 1GB |
| External FLASH | NOR 128 MB and NAND 2 GB |
| ADC | Channel × 6 (12 bits) |
| JTAG | Programming and Debug |
| USB | Debug |
| UART | Monitor |
The orbit height varies for different applications.
The parts in the auxiliary processing unit (Texas Instruments© TMS 570, Dallas, TX, USA).
| Parts | Description |
|---|---|
| Auxiliary processing unit | Texas Instruments TMS 570 |
| External RAM | SRAM 4 MB |
| External FLASH | NAND 1 GB |
| ADC | Channel × 6 (12 bits) |
| JTAG | Programming and Debug |
| USB | Debug |
| UART | Monitor |
The orbit height varies for different applications.
Evaluation indicators: peak signal-to-noise ratio (PSNR), (root mean squared error) RMSE, erreur relative globale adimensionnelle de synthese (ERGAS), and spectral angle mapper (SAM)) for TB-CNN, SSR-NET, SpatCNN, SpecCNN, ConSSFCNN, SSFCNN, ResTFNet, TFNet, and MSDCNN where symbols “↑” or “↓” indicate that the evaluation indicators are directly or inversely proportional to image quality, respectively.
| Scheme | PSNR↑ | RMSE↓ | ERGAS↓ | SAM↓ |
|---|---|---|---|---|
| TBCNN | 37.04 | 0.013 | 1.04 | 3.49 |
| SSRNET | 32.58 | 0.021 | 2.11 | 2.64 |
| SpatCNN | 24.80 | 0.052 | 4.98 | 9.99 |
| SpecCNN | 19.11 | 0.101 | 9.65 | 19.81 |
| ConSSFCNN | 18.73 | 0.105 | 10.35 | 20.72 |
| SSFCNN | 25.81 | 0.047 | 4.43 | 8.49 |
| ResTFNet | 27.99 | 0.042 | 4.11 | 6.33 |
| TFNet | 26.12 | 0.045 | 4.34 | 6.50 |
| MSDCNN | 27.99 | 0.036 | 3.52 | 4.74 |
Urban dataset evaluation metrics for TBCNN, SSRNET, SpatCNN, SpecCNN, ConSSFCNN, SSFCNN, ResTFNet, TFNet, and MSDCNN where symbols “↑” or “↓” indicate that the evaluation indicators are directly or inversely proportional to image quality, respectively.
| Scheme | PSNR↑ | RMSE↓ | ERGAS↓ | SAM↓ |
|---|---|---|---|---|
| TBCNN | 34.69 | 0.018 | 1.16 | 2.46 |
| SSRNET | 28.09 | 0.039 | 5.39 | 4.27 |
| SpatCNN | 23.09 | 0.071 | 6.05 | 7.73 |
| SpecCNN | 18.39 | 0.120 | 7.15 | 13.10 |
| ConSSFCNN | 15.52 | 0.168 | 9.49 | 18.80 |
| SSFCNN | 28.46 | 0.038 | 5.28 | 4.07 |
| ResTFNet | 28.34 | 0.038 | 4.78 | 4.12 |
| TFNET | 28.13 | 0.039 | 5.09 | 4.22 |
| MSDCNN | 28.01 | 0.040 | 5.43 | 4.24 |
Pavia University dataset evaluation metrics for TBCNN, SSRNET, SpatCNN, SpecCNN, ConSSFCNN, SSFCNN, ResTFNet, TFNet, and MSDCNN where symbols “↑” or “↓” indicate that the evaluation indicators are directly or inversely proportional to image quality, respectively.
| Scheme | PSNR↑ | RMSE↓ | ERGAS↓ | SAM↓ |
|---|---|---|---|---|
| TBCNN | 34.05 | 0.011 | 1.65 | 2.76 |
| SSRNET | 29.82 | 0.032 | 5.74 | 5.37 |
| SpatCNN | 27.27 | 0.043 | 10.80 | 10.80 |
| SpecCNN | 33.29 | 0.022 | 4.92 | 6.85 |
| ConSSFCNN | 29.42 | 0.034 | 6.40 | 6.35 |
| SSFCNN | 29.85 | 0.032 | 5.72 | 5.19 |
| ResTFNet | 30.32 | 0.031 | 5.61 | 5.24 |
| TFNET | 30.59 | 0.030 | 5.26 | 4.39 |
| MSDCNN | 29.87 | 0.032 | 5.77 | 5.24 |
Pavia Centre dataset evaluation metrics for TBCNN, SSRNET, SpatCNN, SpecCNN, ConSSFCNN, SSFCNN, ResTFNet, TFNet, and MSDCNN where symbols “↑” or “↓” indicate that the evaluation indicators are directly or inversely proportional to image quality, respectively.
| Scheme | PSNR↑ | RMSE↓ | ERGAS↓ | SAM↓ |
|---|---|---|---|---|
| TBCNN | 35.08 | 0.010 | 1.54 | 2.91 |
| SSRNET | 27.65 | 0.035 | 5.60 | 5.50 |
| SpatCNN | 25.06 | 0.048 | 8.28 | 10.12 |
| SpecCNN | 31.14 | 0.024 | 4.22 | 6.96 |
| ConSSFCNN | 25.10 | 0.047 | 6.88 | 11.14 |
| SSFCNN | 27.47 | 0.036 | 5.69 | 4.93 |
| ResTFNet | 28.58 | 0.032 | 5.22 | 5.54 |
| TFNET | 29.27 | 0.029 | 4.52 | 4.37 |
| MSDCNN | 27.38 | 0.036 | 5.73 | 5.18 |
Botswana dataset evaluation metrics for TBCNN, SSRNET, SpatCNN, SpecCNN, ConSSFCNN, SSFCNN, ResTFNet, TFNet, and MSDCNN where symbols “↑” or “↓” indicate that the evaluation indicators are directly or inversely proportional to image quality, respectively.
| Scheme | PSNR↑ | RMSE↓ | ERGAS↓ | SAM↓ |
|---|---|---|---|---|
| TBCNN | 38.26 | 0.009 | 0.86 | 2.13 |
| SSRNET | 24.61 | 0.047 | 5.21 | 4.06 |
| SpatCNN | 23.83 | 0.052 | 5.81 | 4.80 |
| SpecCNN | 20.38 | 0.077 | 7.33 | 15.95 |
| ConSSFCNN | 20.24 | 0.078 | 7.72 | 13.31 |
| SSFCNN | 24.61 | 0.047 | 5.10 | 3.73 |
| ResTFNet | 25.27 | 0.044 | 5.60 | 4.42 |
| TFNET | 26.55 | 0.038 | 4.41 | 3.77 |
| MSDCNN | 24.52 | 0.048 | 5.14 | 3.81 |
Indian Pines dataset evaluation metrics for TBCNN, SSRNET, SpatCNN, SpecCNN, ConSSFCNN, SSFCNN, ResTFNet, TFNet, and MSDCNN where symbols “↑” or “↓” indicate that the evaluation indicators are directly or inversely proportional to image quality, respectively.
| Scheme | PSNR↑ | RMSE↓ | ERGAS↓ | SAM↓ |
|---|---|---|---|---|
| TBCNN | 36.06 | 0.015 | 1.05 | 1.77 |
| SSRNET | 24.21 | 0.057 | 3.83 | 2.52 |
| SpatCNN | 22.12 | 0.073 | 4.56 | 6.79 |
| SpecCNN | 20.84 | 0.084 | 5.31 | 10.74 |
| ConSSFCNN | 19.51 | 0.098 | 6.66 | 10.64 |
| SSFCNN | 23.79 | 0.060 | 4.06 | 4.51 |
| ResTFNet | 24.53 | 0.055 | 3.72 | 3.63 |
| TFNET | 29.33 | 0.032 | 2.03 | 3.07 |
| MSDCNN | 26.19 | 0.045 | 3.05 | 2.90 |
References
1. Qian, S.-E. Overview of Hyperspectral Imaging Remote Sensing from Satellites. Advances in Hyperspectral Image Processing Techniques; Wiley: Hoboken, NJ, USA, 2022; pp. 41-66.
2. Qian, S.E. Hyperspectral Satellites, Evolution, and Development History. IEEE J. Sel. Top. Appl. Earth Obs. Remote Sens.; 2021; 14, pp. 7032-7056. [DOI: https://dx.doi.org/10.1109/JSTARS.2021.3090256]
3. Stuart, M.B.; McGonigle, A.J.S.; Willmott, J.R. Hyperspectral Imaging in Environmental Monitoring: A Review of Recent Developments and Technological Advances in Compact Field Deployable Systems. Sensors; 2019; 19, 3071. [DOI: https://dx.doi.org/10.3390/s19143071] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/31336796]
4. Stuart, M.B.; Davies, M.; Hobbs, M.J.; Pering, T.D.; McGonigle, A.J.S.; Willmott, J.R. High-Resolution Hyperspectral Imaging Using Low-Cost Components: Application within Environmental Monitoring Scenarios. Sensors; 2022; 22, 4652. [DOI: https://dx.doi.org/10.3390/s22124652]
5. Nolde, M.; Plank, S.; Riedlinger, T. Utilization of Hyperspectral Remote Sensing Imagery for Improving Burnt Area Mapping Accuracy. Remote Sens.; 2021; 13, 5029. [DOI: https://dx.doi.org/10.3390/rs13245029]
6. Krekeler, M.P.S.; Burke, M.; Allen, S.; Sather, B.; Chappell, C.; McLeod, C.L.; Loertscher, C.; Loertscher, S.; Dawson, C.; Brum, J. et al. A novel hyperspectral remote sensing tool for detecting and analyzing human materials in the environment: A geoenvironmental approach to aid in emergency response. Environ. Earth Sci.; 2023; 82, 109. [DOI: https://dx.doi.org/10.1007/s12665-023-10761-1]
7. Ghrefat, H.; Awawdeh, M.; Howari, F.; Al-Rawabdeh, A. Chapter 12—Mineral exploration using multispectral and hyperspectral remote sensing data. Geoinformatics for Geosciences; Stathopoulos, N.; Tsatsaris, A.; Kalogeropoulos, K. Elsevier: Amsterdam, The Netherlands, 2023; pp. 197-222.
8. Pour, A.B.; Zoheir, B.; Pradhan, B.; Hashim, M. Editorial for the Special Issue: Multispectral and Hyperspectral Remote Sensing Data for Mineral Exploration and Environmental Monitoring of Mined Areas. Remote Sens.; 2021; 13, 519. [DOI: https://dx.doi.org/10.3390/rs13030519]
9. Ivliev, N.; Podlipnov, V.; Petrov, M.; Tkachenko, I.; Ivanushkin, M.; Fomchenkov, S.; Markushin, M.; Skidanov, R.; Khanenko, Y.; Nikonorov, A. et al. 3U CubeSat-Based Hyperspectral Remote Sensing by Offner Imaging Hyperspectrometer with Radially-Fastened Primary Elements. Sensors; 2024; 24, 2885. [DOI: https://dx.doi.org/10.3390/s24092885]
10. Soppa, M.A.; Silva, B.; Steinmetz, F.; Keith, D.; Scheffler, D.; Bohn, N.; Bracher, A. Assessment of Polymer Atmospheric Correction Algorithm for Hyperspectral Remote Sensing Imagery over Coastal Waters. Sensors; 2021; 21, 4125. [DOI: https://dx.doi.org/10.3390/s21124125]
11. Govender, M.; Chetty, K.; Bulcock, H. A review of hyperspectral remote sensing and its application in vegetation and water resource studies. Water SA; 2007; 33, pp. 145-151. [DOI: https://dx.doi.org/10.4314/wsa.v33i2.49049]
12. Grøtte, M.E.; Birkeland, R.; Honoré-Livermore, E.; Bakken, S.; Garrett, J.L.; Prentice, E.F.; Sigernes, F.; Orlandić, M.; Gravdahl, J.T.; Johansen, T.A. Ocean Color Hyperspectral Remote Sensing with High Resolution and Low Latency—The HYPSO-1 CubeSat Mission. IEEE Trans. Geosci. Remote Sens.; 2022; 60, pp. 1-19. [DOI: https://dx.doi.org/10.1109/TGRS.2021.3080175]
13. Furano, G.; Tavoularis, A.; Rovatti, M. AI in space: Applications examples and challenges. Proceedings of the 2020 IEEE International Symposium on Defect and Fault Tolerance in VLSI and Nanotechnology Systems (DFT); Frascati, Italy, 19–21 October 2020; pp. 1-6.
14. Wijata, A.M.; Foulon, M.F.; Bobichon, Y.; Vitulli, R.; Celesti, M.; Camarero, R.; Cosimo, G.D.; Gascon, F.; Longépé, N.; Nieke, J. et al. Taking Artificial Intelligence Into Space Through Objective Selection of Hyperspectral Earth Observation Applications: To bring the “brain” close to the “eyes” of satellite missions. IEEE Geosci. Remote Sens. Mag.; 2023; 11, pp. 10-39. [DOI: https://dx.doi.org/10.1109/MGRS.2023.3269979]
15. Corson, M.R.; Korwan, D.R.; Lucke, R.L.; Snyder, W.A.; Davis, C.O. The Hyperspectral Imager for the Coastal Ocean (HICO) on the International Space Station. Proceedings of the IGARSS 2008—2008 IEEE International Geoscience and Remote Sensing Symposium; Boston, MA, USA, 7–11 July 2008; pp. IV-101–IV-104.
16. Puig-Suari, J.; Turner, C.; Ahlgren, W. Development of the standard CubeSat deployer and a CubeSat class PicoSatellite. Proceedings of the 2001 IEEE Aerospace Conference Proceedings (Cat. No.01TH8542); Big Sky, MT, USA, 10–17 March 2001; Volume 341, pp. 1/347-341/353.
17. Villela, T.; Costa, C.A.; Brandão, A.M.; Bueno, F.T.; Leonardi, R. Towards the Thousandth CubeSat: A Statistical Overview. Int. J. Aerosp. Eng.; 2019; 2019, 5063145. [DOI: https://dx.doi.org/10.1155/2019/5063145]
18. Prentice, E.F.; Grøtte, M.E.; Sigernes, F.; Johansen, T.A. Design of a hyperspectral imager using COTS optics for small satellite applications. Proceedings of the International Conference on Space Optics—ICSO 2020; Online, 30 March–2 April 2021; pp. 2154-2171.
19. Whiting, M.; Ustin, S.; Zarco-Tejada, P.; Palacios-Orueta, A.; Vanderbilt, V. Hyperspectral Mapping of Crop and Soils for Precision Agriculture; SPIE: Bellingham, WA, USA, 2006; Volume 6298.
20. Yao, H.L.; Tang, L.; Brown, R.L.; Bhatnagar, D.; Cleveland, T.E. Using Hyperspectral Data in Precision Farming Applications. Advanced Applications in Remote Sensing of Agricultural Crops and Natural Vegetation; CRC Press: Boca Raton, FL, USA, 2011; [DOI: https://dx.doi.org/10.1201/b11222-35]
21. Ennis, R.; Schiller, F.; Toscani, M.; Gegenfurtner, K.R. Hyperspectral database of fruits and vegetables. J. Opt. Soc. Am. A; 2018; 35, pp. B256-B266. [DOI: https://dx.doi.org/10.1364/JOSAA.35.00B256]
22. Ruszczak, B.; Wijata, A.M.; Nalepa, J. Unbiasing the Estimation of Chlorophyll from Hyperspectral Images: A Benchmark Dataset, Validation Procedure and Baseline Results. Remote Sens.; 2022; 14, 5526. [DOI: https://dx.doi.org/10.3390/rs14215526]
23. Jiang, X.; Luo, S.; Ye, Q.; Li, X.; Jiao, W. Hyperspectral Estimates of Soil Moisture Content Incorporating Harmonic Indicators and Machine Learning. Agriculture; 2022; 12, 1188. [DOI: https://dx.doi.org/10.3390/agriculture12081188]
24. Transon, J.; D’Andrimont, R.; Maugnard, A.; Defourny, P. Survey of Hyperspectral Earth Observation Applications from Space in the Sentinel-2 Context. Remote Sens.; 2018; 10, 157. [DOI: https://dx.doi.org/10.3390/rs10020157]
25. Zhu, Q.; Shen, F.; Shang, P.; Pan, Y.; Li, M. Hyperspectral Remote Sensing of Phytoplankton Species Composition Based on Transfer Learning. Remote Sens.; 2019; 11, 2001. [DOI: https://dx.doi.org/10.3390/rs11172001]
26. Christopher, S.; Gupta, P. Global Distribution of Column Satellite Aerosol Optical Depth to Surface PM2.5 Relationships. Remote Sens.; 2020; 12, 1985. [DOI: https://dx.doi.org/10.3390/rs12121985]
27. Salazar-Vazquez, J.; Mendez-Vazquez, A. A plug-and-play Hyperspectral Imaging Sensor using low-cost equipment. HardwareX; 2020; 7, e00087. [DOI: https://dx.doi.org/10.1016/j.ohx.2019.e00087]
28. Esposito, M.; Zuccaro Marchi, A. In-Orbit Demonstration of the First Hyperspectral Imager for Nanosatellites; SPIE: Bellingham, WA, USA, 2019; Volume 11180.
29. Zuccaro Marchi, A.; Maresi, L.; Taccola, M. Technologies and Designs for Small Optical Missions; SPIE: Bellingham, WA, USA, 2019; Volume 11180.
30. Hill, S.L.; Clemens, P. Miniaturization of high spectral spatial resolution hyperspectral imagers on unmanned aerial systems. SPIE Proc.; 2015; 9482, pp. 345-359. [DOI: https://dx.doi.org/10.1117/12.2193706]
31. Praks, J.; Niemelä, P.; Näsilä, A.; Kestilä, A.; Jovanovic, N.; Riwanto, B.; Tikka, T.; Leppinen, H.; Vainio, R.; Janhunen, P. Miniature Spectral Imager in-Orbit Demonstration Results from Aalto-1 Nanosatellite Mission. Proceedings of the IGARSS 2018—2018 IEEE International Geoscience and Remote Sensing Symposium; Valencia, Spain, 22–27 July 2018; pp. 1986-1989.
32. Malan, D.; Palun, A.; Dean, B.; Rotteveel, J.; Stanton, D. Scalable CubeSat Earth Observation payloads, born from international collaboration. Proceedings of the 69th International Astronautical Congress (IAC); Bremen, Germany, 1–5 October 2018.
33. Bender, H.; Mouroulis, P.; Dierssen, H.; Painter, T.; Thompson, D.; Smith, C.; Gross, J.; Green, R.; Haag, J.; Van Gorp, B. et al. Snow and Water Imaging Spectrometer: Mission and instrument concepts for earth-orbiting CubeSats. J. Appl. Remote Sens.; 2018; 12, 044001. [DOI: https://dx.doi.org/10.1117/1.JRS.12.044001]
34. Horiba Grating Catalog. Available online: https://www.horiba.com/int/scientific/products/detail/action/show/Product/spectrometers-cp140-imaging-spectrograph-1754/ (accessed on 25 February 2024).
35. The CMOS sensor Onsemi PYTHON5000. Available online: https://www.onsemi.com/products/sensors/image-sensors/PYTHON5000 (accessed on 25 February 2024).
36. Yuan, Y.; Wu, Y.; Fan, X.; Gong, M.; Ma, W.; Miao, Q. EGST: Enhanced Geometric Structure Transformer for Point Cloud Registration. IEEE Trans. Vis. Comput. Graph.; 2024; 30, pp. 6222-6234. [DOI: https://dx.doi.org/10.1109/TVCG.2023.3329578] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/37971922]
37. Wu, T.; Qian, H.; Liu, Z.; Zhou, J.; Zhou, A. Bi-objective evolutionary Bayesian network structure learning via skeleton constraint. Front. Comput. Sci.; 2023; 17, 176350. [DOI: https://dx.doi.org/10.1007/s11704-023-2740-6]
38. Yang, J.; Zhao, Y.-Q.; Chan, J.C.-W. Hyperspectral and Multispectral Image Fusion via Deep Two-Branches Convolutional Neural Network. Remote Sens.; 2018; 10, 800. [DOI: https://dx.doi.org/10.3390/rs10050800]
39. Zhang, X.; Huang, W.; Wang, Q.; Li, X. SSR-NET: Spatial–Spectral Reconstruction Network for Hyperspectral and Multispectral Image Fusion. IEEE Trans. Geosci. Remote Sens.; 2021; 59, pp. 5953-5965. [DOI: https://dx.doi.org/10.1109/TGRS.2020.3018732]
40. Mei, S.; Yuan, X.; Ji, J.; Zhang, Y.; Wan, S.; Du, Q. Hyperspectral Image Spatial Super-Resolution via 3D Full Convolutional Neural Network. Remote Sens.; 2017; 9, 1139. [DOI: https://dx.doi.org/10.3390/rs9111139]
41. Kingma, D.P.; Ba, J. Adam: A method for stochastic optimization. arXiv; 2014; arXiv: 1412.6980
42. Liu, X.; Liu, Q.; Wang, Y. Remote sensing image fusion based on two-stream fusion network. Inf. Fusion; 2020; 55, pp. 1-15. [DOI: https://dx.doi.org/10.1016/j.inffus.2019.07.010]
43. Han, X.H.; Shi, B.; Zheng, Y. SSF-CNN: Spatial and Spectral Fusion with CNN for Hyperspectral Image Super-Resolution. Proceedings of the 2018 25th IEEE International Conference on Image Processing (ICIP); Athens, Greece, 7–10 October 2018; pp. 2506-2510.
44. Yuan, Q.; Wei, Y.; Meng, X.; Shen, H.; Zhang, L. A Multiscale and Multidepth Convolutional Neural Network for Remote Sensing Imagery Pan-Sharpening. IEEE J. Sel. Top. Appl. Earth Obs. Remote Sens.; 2018; 11, pp. 978-989. [DOI: https://dx.doi.org/10.1109/JSTARS.2018.2794888]
45. Wald, L. Quality of high resolution synthesised images: Is there a simple criterion?. Proceedings of the Third Conference Fusion of Earth Data: Merging Point Measurements, Raster Maps and Remotely Sensed Images; Sophia Antipolis, France, 26–28 January 2000; pp. 99-103.
46. Yuhas, R.H.; Goetz, A.F.; Boardman, J.W. Discrimination among semi-arid landscape endmembers using the spectral angle mapper (SAM) algorithm. Proceedings of the JPL, Summaries of the Third Annual JPL Airborne Geoscience Workshop; Pasadena, CA, USA, 1–5 June 1992; AVIRIS Workshop: Pasadena, CA, USA, 1992; Volume 1.
© 2025 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (https://creativecommons.org/licenses/by/4.0/). Notwithstanding the ProQuest Terms and Conditions, you may use this content in accordance with the terms of the License.
Details
; Chang, Loren C 1
; Chi-Kuang, Chao 1 ; Liu, Jann-Yenq 1
1 Center for Astronautical Physics and Engineering, National Central University, Taoyuan City 320317, Taiwan;
2 Center for Astronautical Physics and Engineering, National Central University, Taoyuan City 320317, Taiwan;
3 Center for Astronautical Physics and Engineering, National Central University, Taoyuan City 320317, Taiwan;