Estimation of Silage Maize Plant Moisture Content Based on UAV Multispectral Data and Ensemble Learning Methods

Abstract

Plant moisture content (PMC) serves as a crucial indicator of crop water status, directly affecting agricultural productivity, product quality, and the effectiveness of precision irrigation. Conventional methods for PMC assessment predominantly rely on destructive sampling techniques, which are labor-intensive and impede real-time monitoring. This study investigates silage maize cultivated in the Hexi region of China, leveraging multispectral data acquired via an unmanned aerial vehicle (UAV) to estimate PMC across different phenological stages. A stacked ensemble learning framework was developed, integrating Back Propagation Neural Network (BPNN), Random Forest Regression (RFR), and Support Vector Regression (SVR), with Partial Least Squares Regression (PLSR) employed for feature fusion. The findings indicate that incorporating vegetation indices into spectral variables significantly improved prediction performance. The standalone models demonstrated coefficient of determination (R²) values ranging from 0.43 to 0.69, with root mean square error (RMSE) spanning 0.61% to 1.43%. In contrast, the ensemble model exhibited superior accuracy, achieving R² values between 0.61 and 0.87 and RMSE values from 0.54% to 1.38%. This methodology offers a scalable, non-invasive alternative for PMC estimation, facilitating data-driven irrigation optimization in regions facing water scarcity.

1. Introduction

Plant moisture content (PMC) is a fundamental parameter for assessing crop water status, playing a crucial role in determining yield, quality, and the effectiveness of precision irrigation strategies in modern agriculture [1]. Reliable and timely estimation of PMC is essential for optimizing water use efficiency and supporting sustainable agricultural practices, especially in response to escalating environmental pressures [2]. While irrigation remains indispensable for crop growth, inefficiencies in water application and excessive usage persist, particularly in arid regions such as the Hexi Corridor in China [3]. These challenges are further compounded by climate change, which exacerbates heat stress and alters precipitation dynamics, necessitating advanced monitoring techniques to mitigate water deficits and enhance adaptive capacity [4]. Traditional PMC measurement methods primarily depend on destructive sampling, requiring plant tissue collection, drying, and weighing to determine moisture content. Although these techniques ensure high accuracy, they are labor-intensive, time-consuming, and unsuitable for large-scale or high-frequency assessments [5]. This underscores the urgent need for innovative, non-invasive approaches to improve monitoring efficiency and facilitate sustainable water resource management [6]. Recent advances in remote sensing and unmanned aerial vehicle (UAV)-based monitoring have shown significant potential for providing real-time PMC data, offering a promising solution for precision irrigation and climate resilience strategies [7].

Conventional PMC assessment predominantly relies on destructive sampling, which, despite its precision, is impractical for large-scale agricultural applications due to extensive labor requirements and low temporal efficiency [8]. In contrast, remote sensing offers a rapid, non-destructive alternative, which can be categorized into three main modalities: satellite-based, UAV-based, and ground-based sensing. Ground-based remote sensing primarily utilizes hyperspectral imaging to assess vegetation water status. Prior research has successfully applied hyperspectral spectroscopy to estimate moisture content in harvested tobacco leaves [9], and recent studies have explored the feasibility of using smartphone-based imaging for tobacco water content assessment [10]. However, when compared to satellite and UAV-based methods, ground-based approaches demonstrate lower efficiency for large-scale crop monitoring. Significant progress has been made in vegetation water content retrieval using satellite remote sensing. Researchers have developed various retrieval models, including regression-based methods that correlate canopy water content with vegetation indices [11], the integration of the PROSAIL radiative transfer model with the Normalized Difference Water Index (NDWI) for estimating moisture levels in low-stature vegetation [12], and the INFORM model for assessing canopy moisture variations in fire-affected landscapes [13]. Additionally, the combination of the Enhanced Vegetation Index (EVI) with NDWI has proven effective in estimating winter wheat canopy water content and analyzing its water balance [14]. Despite these advancements, satellite-based approaches are constrained by long revisit cycles, limiting their temporal resolution. Furthermore, the spatial resolution of freely accessible satellite imagery is often inadequate for precision agriculture, while commercial high-resolution satellite data entail significant costs, restricting their widespread adoption. UAV-based remote sensing mitigates these limitations, offering superior spatial resolution, operational flexibility, and real-time data acquisition, making it increasingly valuable in precision agriculture. UAVs equipped with thermal infrared and multispectral sensors can capture key crop physiological parameters, including canopy temperature and vegetation indices, which can be utilized to estimate leaf area index, chlorophyll content, and water stress levels [15]. Additionally, UAV-mounted hyperspectral cameras, in combination with artificial neural networks, have achieved high-precision estimation of soil moisture content [16]. Other studies have demonstrated that integrating UAV-derived high-temporal-resolution multispectral data with soil moisture observations enables precise estimation of maize biomass and water-use efficiency at the field scale [17]. These findings highlight the significant potential of UAV-based remote sensing in precision agriculture and underscore the advantages of multi-sensor fusion in improving retrieval accuracy. Although hyperspectral and thermal infrared sensors provide higher measurement precision, their substantial cost remains a limiting factor. With the growing adoption of UAV-based multispectral imaging, this technology is expected to emerge as a more cost-effective, rapid, and efficient solution for non-destructive PMC estimation.

The application of machine learning in agricultural remote sensing has expanded significantly, driven by advances in computational techniques and the increasing availability of high-resolution datasets. Machine learning has demonstrated strong potential in tasks such as crop nutrient monitoring and canopy feature analysis, with models like Random Forest (RF) and Gaussian Process Regression (GPR) yielding promising results [18,19]. RF is particularly effective in handling high-dimensional data, identifying key feature variables, and accurately predicting crop nitrogen content [20]. Meanwhile, GPR has been recognized for its robustness in modeling uncertainties, making it a reliable approach for canopy characterization [21]. However, most existing research has focused on single-model applications, with limited exploration of ensemble learning techniques [22]. While individual models can achieve satisfactory results for specific datasets, they often suffer from limitations such as susceptibility to data distribution variations and a propensity for overfitting or underfitting [23]. The integration of multiple machine learning models presents an opportunity to enhance prediction accuracy and improve generalization capabilities in PMC estimation [24].

This study investigates PMC estimation in silage maize cultivated in the Hexi region, leveraging UAV-based multispectral imaging data collected at various growth stages under different irrigation treatments. Three machine learning models—Back Propagation Neural Network (BPNN), Random Forest Regression (RFR), and Support Vector Regression (SVR)—were employed for data fusion, and a stacked ensemble learning framework was implemented to enhance predictive performance. The specific objectives of this research are as follows: (1) to assess the potential of UAV-acquired multispectral imaging data across different growth stages for PMC estimation in silage maize, incorporating spectral bands and vegetation indices; (2) to evaluate the predictive performance of regression models, including BPNN, RFR, and SVR, in estimating PMC; and (3) to develop an ensemble learning approach to improve prediction accuracy and robustness in PMC estimation.

2. Materials and Methods

2.1. Overview of the Study Area and Experimental Design

The experiment was conducted at Huarui Ranch, Minle County, Zhangye City, Gansu Province, China (100°22′–101°13′ E, 37°56′–38°48′ N). This area is located within the Heihe River system and experiences a temperate continental desert–steppe climate [25]. The geographic location of the study area and plant sampling points are shown in Figure 1. The long-term average sunshine duration is 2794 h per year, with an average annual precipitation of 351 mm and an average annual temperature of 6.5 °C. During the 2022 silage maize growing season, the total precipitation was 181.4 mm, with a frost-free period of 140 days. To optimize water use, drip irrigation and sprinkler irrigation are commonly employed, with drip irrigation being the predominant method in silage maize cultivation. Crop rotation typically involves alternating maize with wheat or barley to maintain soil fertility and reduce pest pressure. In addition to silage maize, the region’s major crops include oats and alfalfa, with crop selection largely influenced by water availability and market demand. Climatic conditions during the silage maize growing season are shown in Figure 2.

The experiment was conducted from April to October 2022. Stratification was conducted based on soil texture, topographic gradient, and irrigation pattern. Given the dominance of sandy soil, minimal elevation variation, and proximity to drip irrigation pipelines, the sampling method followed an “S”-shaped five-point sampling approach. Sampling points were distributed throughout the field to capture spatial variability. The experiment consisted of four irrigation treatments: W1 (30% water-saving compared to the traditional irrigation rate, 4305 m³/hm²), W2 (20% water-saving, 4920 m³/hm²), W3 (10% water-saving, 5535 m³/hm²), and W4 (traditional drip irrigation rate, 6150 m³/hm²). Each treatment was replicated three times, resulting in a total of 12 experimental plots. Each plot measured 25 m in length and 1.6 m in width, covering an area of 40 m². To minimize interference between adjacent plots, a 1 m buffer zone was established between them.

The maize variety used in the experiment was silage “Jinling No. 67”, provided by Silage Maize Seed Company (Gansu Huarui Agriculture Co., Ltd., Zhangye, China). The planting distance was 25 cm between plants and 50 cm between rows. Irrigation was carried out using subsurface drip irrigation, with tape emitters of the built-in patch type. The diameter of the laterals was 16 mm, with a wall thickness of 0.2 mm, and the distance between emitters was 300 mm, with a discharge rate of 2.2 L/h. The drip tape was laid parallel to the direction of silage maize rows, spaced 60 cm apart. Fertilization was performed using a hydraulic pump system, applied during irrigation. The nitrogen fertilizer used was urea (CO(NH₂)₂, N 46%), with a total application rate of 311 kg/hm². Phosphorus and potassium fertilizers were applied as diammonium phosphate (containing 44% P₂O₅) and potassium sulfate (containing 57% K₂O), respectively, with nitrogen fertilizer. They were applied in three stages: pre-jointing (5.9%), jointing stage (64.7%), and tasseling stage (29.4%). The total amounts applied were 146 kg/hm² for phosphorus and 107 kg/hm² for potassium.

2.2. Data Collection and Processing

2.2.1. UAV Data Acquisition

In this study, multispectral remote sensing data were collected using a DJI M300 RTK quadcopter (DJI Technology Co., Ltd., Shenzhen, China) equipped with an MS600 Pro multispectral imaging system (Changguang Yuchen Information Technology Equipment Co., Ltd., Qingdao, China). This system comprises six independent CMOS image sensors, each featuring a resolution of 1.2 megapixels and configured to capture data across six distinct spectral bands: blue (450 nm @ 35 nm), green (555 nm @ 27 nm), red (660 nm @ 22 nm), red-edge 1 (720 nm @ 10 nm), red-edge 2 (750 nm @ 10 nm), and near-infrared (840 nm @ 30 nm). The data acquisition process was carried out under clear sky conditions between 10:00 AM and 12:00 PM using a structured vertical aerial imaging strategy. A preprogrammed flight route was executed to ensure systematic coverage of the study site, with radiometric calibration performed prior to data collection utilizing a standardized white reference panel. The UAV operated at a flight altitude of 30 m with a velocity of 2.5 m/s, yielding a ground sampling distance (GSD) of 2 cm. To ensure comprehensive scene coverage and minimize data gaps, the front and side overlap ratios were set to 80% and 75%, respectively. Figure 3 presents an overview of the UAV-mounted multispectral sensor, the calibration panel, and the predefined flight path, while the detailed UAV operational parameters are summarized in Table 1.

2.2.2. Plant Moisture Content Measurement

The study area was systematically divided into 12 sections, and sample collection was conducted using the “S” five-point sampling technique. On cloud-free days corresponding to the silking stage (16 July), tasseling stage (6 August), and physiological maturity stage (15 September) of silage maize, plant specimens exhibiting uniform growth characteristics were gathered from a 5 cm radius around designated sampling locations, concurrent with UAV-based multispectral data acquisition. Each collected plant was immediately placed into a sealed bag, and its fresh weight was measured and documented. The samples underwent an initial drying phase in an oven set at 105 °C for 30 min to remove surface moisture, followed by a secondary drying process at 70 °C for approximately 48 h until a constant mass was attained. Once fully dried, the final dry weight was recorded. The PMC was subsequently determined using the following equation, expressed in terms of wet weight:

$PMC={\displaystyle \frac{FW-DW}{DW}}\times 100$

where PMC is the plant moisture content (%), FW is the fresh weight (mg), and DW is the dry weight (mg).

2.2.3. UAV Image Preprocessing

Four ground control points (GCPs) were strategically positioned within the study site, with their precise coordinates determined using real-time kinematic (RTK) positioning technology. Image alignment and georeferencing were conducted using Pix4D Mapper 4.8.0, incorporating manually marked control points to enhance positional accuracy [26]. The final orthomosaic achieved a root mean square error (RMSE) of 0.05 m along the x-axis, 0.05 m along the y-axis, and 0.10 m along the z-axis, demonstrating a high degree of spatial precision. During UAV data acquisition, variations in solar illumination angles and sensor viewing geometry introduced shadow effects, which compromised the spectral fidelity of crop canopy reflectance. To mitigate these artifacts, supervised classification was performed in ENVI 5.3 using false-color composite imagery. The maximum likelihood classification method was applied to eliminate background soil interference and shadowed areas, followed by radiometric calibration [27]. Digital number (DN) values were converted into reflectance using pre-calibrated imagery with established reflectance references, ensuring spectral consistency across all bands. Additionally, manufacturer-provided calibration coefficients for the MS600 Pro multispectral sensor (Changguang Yuchen Information Technology Equipment Co., Ltd., Qingdao, China) were utilized to compensate for spectral response variations across six bands: blue, green, red, red-edge 1, red-edge 2, and near-infrared [28]. A standard reflectance panel was employed for in-field radiometric calibration to account for illumination fluctuations during UAV flight missions. To normalize lighting inconsistencies across different phenological stages, histogram matching was implemented, aligning spectral characteristics within the same growth phase. In ArcMap 10.8, a Shapefile was created to georeference plant sampling points, incorporating 60 uniquely labeled locations from which spectral data were extracted. The calculated PMC values were spatially integrated with corresponding GPS coordinates and converted into a point dataset, subsequently overlaid onto the UAV-derived multispectral imagery of the study area. Spectral indices were computed using the raster calculator in ArcMap 10.8, leveraging multispectral band data, and vegetation indices were retrieved based on the unique ID associated with each sampling site [29]. To minimize edge distortions in the imagery, a study area boundary mask was generated, and the dataset was cropped to retain only the designated region of interest. The distribution of sampling points and control points are shown in Figure 4.

2.3. Spectral Index Construction

To collect relevant spectral vegetation indices (VIs) for this study, an index database was referenced [30] Based on the spectral range of the UAV multispectral images, 14 spectral indices that have been proven to be highly correlated with plant moisture content in the literature were selected, as shown in Table 2.

2.4. Model Development

To enhance the precision of PMC predictions, this study introduces a stacking-based ensemble learning framework comprising two primary phases. Initially, three independent PMC prediction models were developed utilizing multispectral data processed through distinct machine learning algorithms: Back Propagation Neural Network (BPNN), Random Forest Regression (RFR), and Support Vector Regression (SVR). Subsequently, predictions from these models were integrated using a Partial Least Squares Regression (PLSR) meta-learner. The selection of these three machine learning approaches is supported by extensive literature validating their efficacy in PMC estimation [39]. Each model contributes unique predictive insights, which are essential for constructing a robust ensemble learning system.

BPNN, a widely used artificial neural network, refines weight parameters iteratively through a backpropagation algorithm to minimize discrepancies between predicted and actual values [40]. Recent studies highlight its effectiveness in agricultural applications, including UAV-based crop yield estimation and hyperspectral remote sensing for soil and vegetation analysis [40,41]. RFR, a decision-tree-based ensemble method, models relationships between input and output variables through hierarchical decision rules. It excels in handling high-dimensional data, assessing variable importance, mitigating overfitting, and improving generalization performance [42]. SVR, on the other hand, performs regression by determining an optimal hyperplane in the feature space that minimizes the cumulative deviation of sample points from the hyperplane, thereby reducing overall prediction errors [43]. PLSR, serving as the meta-learner, integrates principal component analysis, correlation analysis, and multiple linear regression to facilitate multivariate statistical modeling. Its capability to manage collinearity and handle cases where the number of predictor variables exceeds sample size makes it particularly suitable for this study [44].

Stacked regression, an advanced ensemble learning methodology, enhances predictive accuracy by combining multiple base models to leverage complementary data representations [45]. As depicted in Figure 5, the stacking procedure was systematically implemented. The dataset was partitioned using a 5-fold cross-validation strategy, repeated 40 times to ensure model reliability. An out-of-fold (OOF) prediction approach was adopted, wherein the base learners (BPNN, RFR, SVR) were trained on four data folds while making predictions on the remaining fold. This iterative process generated OOF predictions for the entire training dataset, which were subsequently used as meta-features for PLSR training [46]. For the test dataset, meta-features were derived by applying base models—trained on the entire training dataset—to generate test predictions. Each base learner produced five sets of validation predictions corresponding to different training subsets, which were vertically concatenated to construct a test set prediction matrix. This matrix then served as input for the meta-learner. The predictive performance of individual base models was evaluated by averaging results across validation sets. Ultimately, PLSR was trained using the same 5-fold cross-validation scheme to integrate predictions from the base models [47]. The final prediction accuracy was determined by averaging validation results obtained from the test set prediction matrix.

The PMC inversion model was implemented in MATLAB 2022a. Hyperparameter optimization for BPNN, RFR, and SVR was conducted via a grid search strategy combined with 5-fold cross-validation, selecting optimal parameters based on the highest mean R² value. Specifically, BPNN was structured with two hidden layers, each comprising 10 neurons. RFR was configured with 100 decision trees and a maximum depth of 20. For SVR, a radial basis function (RBF) kernel was employed, with regularization parameter C = 1.0 and tolerance ε = 0.1.

2.5. Model Accuracy Evaluation Parameters

The original dataset was partitioned into training and validation subsets, with the procedure repeated 40 times to enhance robustness. A five-fold cross-validation strategy was implemented for model training. Over the course of 40 iterations, a total of 200 test results were generated. The final model performance was assessed based on the average of these test outcomes, utilizing key evaluation metrics: the coefficient of determination (R²), root mean square error (RMSE), and the ratio of performance to deviation (RPD). RMSE quantifies the deviation between predicted and observed values, with lower values indicating greater predictive accuracy. The coefficient of determination (R²) measures the model’s goodness of fit, where values approaching 1 signify superior inversion performance. RPD, widely applied for evaluating agreement between predicted and actual values, serves as an indicator of predictive reliability. Specifically, an RPD exceeding 2 suggests strong predictive capability, values ranging from 1.5 to 2.0 indicate moderate performance, whereas an RPD below 1.5 denotes weak predictive power [48]. The mathematical expressions for these evaluation metrics are as follows:

(1)$R^2=1-{\displaystyle \frac{{\sum }_{i=1}^n{(\widehat{y_i}-y_i)}^2}{{\sum }_{i=1}^n{(y_i-\overline{y_i})}^2}}$

(2)$RMSE=\sqrt{{\displaystyle \frac{{\sum }_{i=1}^n{(y_i-\widehat{y_i})}^2}{N}}}$

(3)$RPD={\displaystyle \frac{SD}{RMSE}}$

where $y_i$ is the actual PMC value, $\widehat{y_i}$ is the predicted PMC value, $\overline{y_i}$ is the mean PMC, N is the sample size, and SD is the standard deviation of the measurements in the prediction set.

3. Results and Analysis

3.1. Correlation Analysis Between Plant Moisture Content and Spectral Indices

Environmental factors such as water stress, nutrient availability, and intrinsic growth characteristics can lead to variations in plant traits, resulting in differences in spectral reflectance properties within the same species. To account for these variations, this study selected plant samples from three distinct growth stages: tasseling, silk emergence, and maturity, with each stage comprising 60 samples. The statistical summary of PMC measurements across these growth stages is presented in Table 3. The analysis of range, standard deviation, and coefficient of variation—both for the overall dataset and for individual growth stages—reveals significant differences in PMC, indicating effective data stratification. As depicted in Figure 6, plant moisture content exhibits a declining trend throughout the growth stages, with all samples closely adhering to a normal distribution. To further examine the relationship between spectral indices and plant moisture content, a Pearson correlation analysis was performed, and the results are visualized in Figure 7. The spectral indices across different growth stages generally met statistical significance criteria, with certain variables exhibiting highly significant correlations (p < 0.001). These findings confirm that the selected spectral indices are suitable for predictive modeling.

3.2. Comprehensive Evaluation of the Models

Three standalone machine learning models, along with one ensemble learning approach, were implemented to predict PMC from multispectral imagery of silage maize. The predictive performance of these models is summarized in Table 4. Among the individual models, Random Forest Regression (RFR) demonstrated the highest accuracy during the tasseling stage when using spectral bands as input features, achieving an R² of 0.54, an RMSE of 1.19%, and an RPD of 1.49. When vegetation indices were incorporated as input variables at the same growth stage, RFR again outperformed other models, yielding an R² of 0.60, an RMSE of 1.17%, and an RPD of 1.70. To assess the impact of vegetation indices on predictive accuracy, their inclusion was evaluated across all four machine learning models. Results indicated that the addition of vegetation indices improved the performance of all models, with Support Vector Regression (SVR) exhibiting the most notable increase in R². Additionally, predictions from the three individual models were integrated using the stacking-based Partial Least Squares Regression (PLSR) ensemble approach, which consistently outperformed the standalone models when using the same input features.

Figure 8 illustrates the observed versus predicted PMC values for the optimal prediction model. During the tasseling stage, utilizing spectral bands alone yielded an R² of 0.66, which improved to 0.87 upon incorporating vegetation indices. In the silk emergence stage, the R² increased from 0.65 to 0.85, while in the maturity stage, it rose from 0.61 to 0.75 following the integration of vegetation indices. These results confirm that the ensemble learning model constructed with both spectral bands and vegetation indices achieved superior predictive accuracy, particularly excelling during the tasseling stage. Notably, RFR contributed the highest model weight within the ensemble, underscoring the significance of high-accuracy individual models in enhancing ensemble performance and reinforcing the interdependence between base and ensemble model effectiveness.

3.3. Spatial Distribution Map of Plant Moisture Content

In this study, the optimal model was employed to estimate the spatial distribution of PMC in silage maize, providing insights into its variations across different growth stages. The results revealed a progressive decline in PMC throughout the growth cycle, primarily driven by water consumption during maize development and the physiological redistribution of water in the maturity stage. As the plants advanced toward maturity, moisture content decreased steadily, leading to an overall reduction in PMC. The spatial distribution map presented in Figure 9 highlights significant heterogeneity in moisture content across the maize field. Notably, higher moisture levels were observed in the central regions, whereas a gradual decline was evident toward the field edges. This spatial variability can be attributed to multiple factors, including soil moisture redistribution, variations in agronomic management, and edge effects. The central areas of the field maintained relatively stable moisture conditions due to reduced exposure to external environmental fluctuations, facilitating higher plant moisture retention. Conversely, peripheral regions exhibited a marked decline in moisture content, likely influenced by factors such as increased wind exposure, elevated evaporation rates, and insufficient soil moisture replenishment. Additionally, disparities in irrigation practices and field management strategies may have contributed to localized variations in moisture availability, further accentuating spatial inconsistencies.

4. Discussion

4.1. Analysis of Spectral Indices

The sensitivity of PMC to various spectral bands and vegetation indices differs across growth stages, reflecting its close association with plant transpiration, water uptake efficiency, and cellular osmotic regulation [49]. During the early developmental phases, particularly the tasseling stage, maize exhibits a high transpiration rate due to extensive stomatal opening and intensified metabolic activity. This leads to pronounced reflectance variations in the near-infrared and red-edge spectral bands, which are highly responsive to leaf water content and chlorophyll concentration [50]. As the crop matures, PMC progressively declines due to senescence and reduced stomatal conductance, subsequently altering spectral reflectance properties. The most substantial reduction in PMC occurs during the maturation stage, aligning with prior research that has documented similar trends in wheat and rice [51]. Overall, fluctuations in PMC are predominantly influenced by chlorophyll activity and the leaf area index, both of which modulate light absorption and reflection in the visible spectrum [52]. This pattern has also been corroborated by Deng et al. [53] and Harris et al. [54], who demonstrated that water stress-induced variations in green leaf reflectance within the near-infrared region render this wavelength a crucial predictor of PMC. Furthermore, our findings underscore the importance of vegetation indices in PMC estimation. Among these, the Normalized Difference Vegetation Index (NDVI) and the Red-Edge Normalized Difference Vegetation Index (RENDVI) exhibited the highest correlations with PMC across different growth stages, reaffirming the sensitivity of the near-infrared and red-edge bands to plant water content [55]. The enhanced predictive accuracy of the integrated model highlights the efficacy of incorporating multiple spectral indices to better capture the nonlinear interactions between spectral reflectance and plant water dynamics [56]. These results further validate the potential of UAV-based remote sensing in precision agriculture, emphasizing the integration of multispectral data with advanced machine learning techniques for improved PMC assessment.

4.2. The Importance of Machine Learning Models in PMC Prediction

The novel aspect of this study lies in the application of bootstrap resampling and stacked ensemble learning models, offering a new approach for PMC prediction with limited sample sizes. The results indicate that integrating spectral bands and vegetation indices with ensemble learning methods significantly enhances the prediction accuracy of PMC in silage maize. The stacked ensemble learning approach achieved the highest predictive performance (R² = 0.87, RMSE = 1.04%), outperforming individual machine learning models. This finding aligns with previous studies emphasizing the effectiveness of ensemble models in agricultural applications [57]. While individual models such as RFR and SVR perform well at specific growth stages, their predictive capabilities fluctuate depending on the dataset used. The ensemble approach effectively balances these inconsistencies by leveraging the strengths of multiple models, leading to more stable and accurate predictions [58]. Furthermore, this study provides insights into model selection for PMC prediction. The RFR model contributed the highest weight in the ensemble, indicating its superior capability in extracting key spectral features from high-dimensional data. This observation is consistent with prior findings in crop phenotyping, where RFR has been demonstrated to outperform other algorithms in handling spectral heterogeneity [59]. However, the combination of different learning algorithms further enhanced prediction robustness, suggesting that future PMC models should integrate multiple data sources and machine learning strategies to maximize accuracy [60]. Previous studies have explored the relationship between spectral indices and PMC, with many relying on single regression models such as RF, BPNN, and SVR [39,43,61]. While these models provide reasonable accuracy, they often suffer from overfitting or limited generalizability. Our study confirms that stacking multiple machine learning models can effectively mitigate these issues, providing more robust predictions across different growth stages. Compared to traditional regression methods, our approach, which integrates spectral features with vegetation indices, significantly improves prediction accuracy, particularly during the tasseling stage. This finding corroborates previous remote sensing studies on crop water stress [62].

4.3. Spatial Variability and Its Implications for Precision Agriculture

The spatial distribution analysis revealed significant heterogeneity in the PMC across the maize field, with higher water content observed in the central region and lower levels near the field edges. This pattern may be attributed to microclimatic variations, soil moisture heterogeneity, and differences in irrigation distribution [63]. Similar findings have been reported in studies on crop water stress, emphasizing the role of environmental factors in shaping water distribution patterns [60]. From an applied perspective, these results underscore the potential of UAV-based multispectral imaging for precision irrigation management. By identifying water-deficient areas, irrigation schedules can be optimized, ensuring precise water application while minimizing waste and maintaining crop health. This approach aligns with recent advancements in high-resolution remote sensing technologies developed for precision agriculture [64]. Furthermore, the findings of this study can be integrated with automated irrigation systems to optimize water allocation based on PMC estimations, thereby reducing water wastage and improving crop yield. The high accuracy of PMC predictions also enhances nutrient management strategies, ensuring that fertilization practices are tailored to the actual water status of plants rather than being based on empirical estimations [50].

4.4. Uncertainty Analysis and Future Directions

Despite the encouraging results of this study, certain limitations must be acknowledged. First, the dataset was collected from a single growing season and location, which may restrict the generalizability of the findings across different environmental and temporal conditions. Future research should integrate datasets from multiple seasons and regions to further validate the robustness of the proposed model [65]. Second, although this study utilized multispectral data, incorporating hyperspectral and thermal infrared imaging could further enhance the accuracy of PMC prediction. Hyperspectral sensors enable the capture of finer spectral characteristics, whereas thermal infrared data provide direct assessments of plant water stress [64]. Moreover, environmental conditions and agricultural management practices may also influence reflectance-based predictions. For instance, high temperatures or drought conditions may induce stress-related chlorophyll variations, consequently altering leaf reflectance. Strong winds can modify canopy structure, affect leaf angle distribution, and introduce noise into spectral data, ultimately reducing the accuracy of spectral indices. Additionally, rainfall before UAV flights may alter surface moisture levels, leading to fluctuations in reflectance values. Although data collection in this study was conducted under stable conditions to minimize these effects, future research should integrate real-time weather data and optimize flight scheduling strategies to improve model robustness and assess the impact of various irrigation and fertilization practices. Another limitation of this study is its dependence on machine learning models that require large training datasets. Although the bootstrap resampling method partially mitigated the small-sample effect, expanding the dataset through additional field measurements would enhance model stability. Future studies should investigate deep learning approaches, including convolutional neural networks (CNNs) and transformers, which have shown remarkable potential in remote sensing applications [66].

5. Conclusions

This study utilized UAV-based remote sensing to assess the predictive accuracy of BPNN, RFR, and SVR models under different input feature configurations, integrating spectral bands and vegetation indices. The predictive performance of these machine learning approaches for estimating PMC in silage maize was systematically evaluated, leading to the following key conclusions:

(1) Effect of Input Features on Model Accuracy: Models relying solely on individual spectral band reflectance exhibited lower predictive accuracy, whereas the incorporation of vegetation indices significantly improved performance. In particular, vegetation indices derived from the near-infrared and red-edge spectral regions demonstrated strong correlations with PMC, highlighting their relevance for predictive modeling.

(2) Performance of Individual Machine Learning Models: When both spectral bands and vegetation indices were used as input features, the individual machine learning models achieved the highest inversion accuracy during the tasseling stage. Among these models, RFR consistently outperformed BPNN and SVR, demonstrating superior predictive accuracy across different growth stages.

(3) Advantage of Ensemble Learning: The ensemble machine learning model, constructed by integrating spectral bands and vegetation indices, outperformed the individual models in predictive accuracy and generalization capability. The ensemble approach achieved a maximum R² of 0.87 and the lowest RMSE of 0.54%, confirming its effectiveness in enhancing PMC estimation.

Author Contributions

Writing—original draft preparation, X.L.; project administration, J.Y. and C.H.; writing—review and editing, J.L. and X.Y.; supervision, W.M. and Z.G.; software, X.L., Q.D., H.Y. and K.C. All authors have read and agreed to the published version of the manuscript.

Data Availability Statement

All data used during the study are proprietary or confidential and only limited data can be provided.

Acknowledgments

We sincerely thank the editor and reviewers for their time and effort in reviewing our work.

Conflicts of Interest

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.

Figures and Tables

Figure 1. (a,b) The study area is located in Minle Ecological Industrial Park, Zhangye City, Gansu Province, China. (c) Photos of silage maize cultivars.

Figure 2. Meteorological data during the 2022 silage maize growing season. (a) Daily average temperature, (b) Daily average radiation, (c) Daily average precipitation.

Figure 3. (a) DJI M300 RTK UAV equipped with MS600 Pro multi-spectral camera; (b) the calibrated reflection panel; (c) flight path planning of the study area.

Figure 4. Sampling point distribution. Note: four ground control points (GCPs) were strategically established within the study area, ensuring an even distribution of sample points across the region.

Figure 5. Ensemble learning model framework. BPNN Back Propagation Neural Network, RFR random forest regression, SVR Support Vector Regression p predictions from different models.

Figure 6. Statistical results of Plant Moisture Content (PMC) at different growth stages.

Figure 7. Pearson correlation analysis. Note: * p ≤ 0.05 ** p ≤ 0.01 *** p ≤ 0.001.

Figure 8. Observed and predicted values of the best PMC prediction model. Note: VI vegetation indices.

Figure 9. Spatial distribution of plant water content.

Table 1

DJI M300 RTK UAV and MS600 Pro multi-spectral camera specifications.

Parameter	Specification
UAV type	Rotary wing
Size	0.81 (width) × 0.67 (length) × 0.43 (height) m
Flight duration	55.00 min
Maximum speed	17.00 m/s
Maximum altitude	7.00 km
Maximum take-off weight	9.00 kg
Maximum flight range	15.00 km
Payload capacity	0.93 kg
MS600 Pro camera weight	0.66 kg

Table 2

List of vegetation indices (VIs) used in the modelling of crop water content.

Index	Full Name	Formula	References
COSRI	Combine spectroscopy index	(B + G)/(R + NIR) × NDVI	[31]
CVI	Chlorophyll vegetation index	(NIR × R)/G2	[32]
DVI	Difference vegetation index	NIR − R	[33]
GNDVI	Green normalized difference vegetation index	(N − G)/(N + G)	[34]
NDREI	Normalized difference red edge index	(N − RE1)/(N + RE1)	[35]
NDVI	Normalized difference vegetation index	(NIR − R)/(NIR + R)	[36]
NGRDI	Normalized difference green/red index	(G − R)/(G + R)	[37]
RENDVI	Red edge normalized difference vegetation index	(RE2 − RE1)/(RE2 + RE1)	[38]
Rblue	Blue	B	/
Rgreen	Green	G	/
Rnir	Nir	N	/
RRE1	RedEdge720	RE1	/
RRE2	RedEdge750	RE2	/
Rred	Red	R	/

Note: B, G, R, RE1, RE2, and NIR are the spectral reflectance of MS600 Pro multispectral camera at wavelengths of 450, 555, 660, 720, 750 and 840 nm, respectively.

Table 3

Statistics on PMC characteristics at each fertility stage.

Category	Observations	Min	Max	Mean	SD	R	CV
All datasets	180	0.65	0.83	0.74	0.06	0.177	0.075
Tasseling stage	60	0.75	0.83	0.78	0.02	0.072	0.020
Silking period	60	0.73	0.78	0.77	0.01	0.053	0.014
Maturity	60	0.65	0.69	0.66	0.02	0.040	0.027

Note: Min minimum, Max maximum, SD standard deviation, R Range, CV coefficient of variation.

Table 4

PMC prediction accuracy based on different machine learning methods.

Growth Stages	Feature Type	Metrics	BPNN	RFR	SVR	Stacking (PLSR)
Tasseling stage	Band	R2	0.50	0.54	0.51	0.66
		RMSE (%)	1.30	1.19	1.27	1.30
		RPD	1.45	1.49	1.59	1.54
	VI	R2	0.54	0.60	0.59	0.78
		RMSE (%)	1.33	1.17	1.29	1.24
		RPD	1.50	1.70	1.55	1.61
	Band + VI	R2	0.63	0.69	0.67	0.87
		RMSE (%)	1.08	1.12	1.05	1.04
		RPD	1.87	1.79	1.90	1.93
Silking period	Band	R2	0.46	0.53	0.49	0.65
		RMSE (%)	0.63	0.68	0.67	0.66
		RPD	1.59	1.47	1.48	1.51
	VI	R2	0.49	0.54	0.52	0.76
		RMSE (%)	0.66	0.65	0.67	0.60
		RPD	1.51	1.52	1.49	1.67
	Band + VI	R2	0.59	0.62	0.61	0.85
		RMSE (%)	0.62	0.61	0.62	0.54
		RPD	1.62	1.62	1.61	1.85
Maturity	Band	R2	0.43	0.49	0.46	0.61
		RMSE (%)	1.35	1.43	1.34	1.38
		RPD	1.48	1.40	1.49	1.45
	VI	R2	0.53	0.57	0.56	0.64
		RMSE (%)	1.24	1.24	1.08	1.29
		RPD	1.61	1.18	1.30	1.55
	Band + VI	R2	0.61	0.65	0.63	0.75
		RMSE (%)	1.24	1.18	1.19	1.15
		RPD	1.61	1.69	1.68	1.74

References

1. Ali, A.; Kaul, H.P. Monitoring Yield and Quality of Forages and Grassland in the View of Precision Agriculture Applications—A Review. Remote Sens.; 2025; 17, 279. [DOI: https://dx.doi.org/10.3390/rs17020279]

2. Cai, X.; Zhang, X.; Noël, P.H.; Shafiee-Jood, M. Impacts of climate change on agricultural water management: A review. Agriculture; 2015; 2, pp. 439-455.

3. Fakhar, M.S.; Nazari, B. Estimating Soil Moisture Using Vegetation Cover Indices and Soil Surface Temperature in Agricultural and Saline Fields of Qazvin Plain. Water Soil Manag. Model.; 2024; [DOI: https://dx.doi.org/10.22098/mmws.2024.14569.1416]

4. Fan, X.; Zhang, H.; Zhou, L.; Bian, L.; Jin, X.; Tang, L.; Ge, Y. Evaluating Drought Stress Response of Poplar Seedlings Using a Proximal Sensing Platform via Multi-Parameter Phenotyping and Two-Stage Machine Learning. Comput. Electron. Agric.; 2024; 225, 109261.

5. Singla, M.; Gupta, D.; Goyal, D.; Dang, R.K. Integrating Smart Farming with AI Techniques for Crop Disease Prediction. Proceedings of the 2024 15th International Conference on Computing Communication and Networking Technologies (ICCCNT); Kamand, India, 24–28 June 2024; pp. 1-6.

6. Mandal, D.; Yilma, W.; Atreya, M.; Kauzya, J.B.; Smock, N.; Khosla, R.; Whiting, G.L. Evaluation of Hybrid Biodegradable Sensor Node for Monitoring Soil Moisture. Harnessing Data Science for Sustainable Agriculture and Natural Resource Management; Springer Nature: Singapore, 2024; pp. 271-298.

7. Zhang, S.; Zhang, C.; Yang, C.; Liu, B. Artificial Intelligence and Internet of Things for Smart Agriculture. Front. Plant Sci.; 2025; 14, 1494279.

8. Franz, A.; Sowiński, J.; Głogowski, A.; Fiałkiewicz, W. A Preliminary Study on the Use of Remote Sensing Techniques to Determine the Nutritional Status and Productivity of Oats on Spatially Variable Sandy Soils. Agronomy; 2025; 15, 616. [DOI: https://dx.doi.org/10.3390/agronomy15030616]

9. Liu, H.; Wu, X.; Li, D.; Zhang, F.; Zhang, D.; Huang, H. Study on the Moisture Content of Tobacco Leaves during the Picking Period Based on Hyperspectral Technology. J. Chin. Agric. Mech.; 2021; 42, pp. 157-163.

10. Zhang, J.; Song, Z.; Han, W.; Shang, X.; Liu, Z. Nondestructive Testing of Tobacco Leaf Water Status by Using Digital RGB Images. J. Chin. Agric. Mech.; 2019; 40, pp. 62-68.

11. Sun, G.; Xing, T.; Xing, Y.; Huang, J. Inversion of Vegetation Canopy Water Content in Jingyuetan Park Based on Remote Sensing Data. For. Eng.; 2019; 35, pp. 32-38.

12. Jiang, H.; Chai, L.; Jia, K.; Liu, J.; Yang, S.; Zheng, J. Estimation of Water Content for Short Vegetation Based on PROSAIL Model and Vegetation Water Indices. Natl. Remote Sens. Bull.; 2021; 25, pp. 1025-1036.

13. Rao, Y.; Liu, K.; Huang, H.; Lin, Q.; Wang, C. Inversion of Canopy Water Content of Fire Scars in Genhe County Based on INFORM Model and Google Earth Engine. J. Cent. South Univ. For. Technol.; 2021; 41, pp. 80-91.

14. Cheng, X.-J.; Yang, G.-J.; Xu, X.-G.; Chen, T.-E.; Li, Z.-H.; Feng, H.-K.; Wang, D. Estimating Canopy Water Content in Wheat Based on New Vegetation Water Index. Spectrosc. Spect. Anal.; 2014; 34, pp. 3391-3396.

15. Berni, J.A.J.; Zarco-Tejada, P.J.; Suárez, L.; Fereres, E. Thermal and Narrowband Multispectral Remote Sensing for Vegetation Monitoring from an Unmanned Aerial Vehicle. IEEE Trans. Geosci. Remote Sens.; 2009; 47, pp. 722-738.

16. Hassan-Esfahani, L.; Torres-Rua, A.; Jensen, A.; McKee, M. Assessment of Surface Soil Moisture Using High-Resolution Multi-Spectral Imagery and Artificial Neural Networks. Remote Sens.; 2015; 7, pp. 2627-2646. [DOI: https://dx.doi.org/10.3390/rs70302627]

17. Han, W.; Tang, J.; Zhang, L.; Niu, Y.; Wang, T. Maize Water Use Efficiency and Biomass Estimation Based on Unmanned Aerial Vehicle Remote Sensing. Trans. Chin. Soc. Agric. Mach.; 2021; 52, pp. 129-141.

18. Chlingaryan, A.; Sukkarieh, S.; Whelan, B. Machine Learning Approaches for Crop Yield Prediction and Nitrogen Status Estimation in Precision Agriculture: A Review. Comput. Electron. Agric.; 2018; 151, pp. 61-69.

19. Akkem, Y.; Biswas, S.K.; Varanasi, A. Smart Farming Using Artificial Intelligence: A Review. Eng. Appl. Artif. Intell.; 2023; 120, 105899.

20. Reddy, D.J.; Kumar, M.R. Crop Yield Prediction Using Machine Learning Algorithm. Proceedings of the 2021 5th International Conference on Intelligent Computing and Control Systems (ICICCS); Madurai, India, 6–8 May 2021; pp. 1466-1470.

21. Hamrani, A.; Akbarzadeh, A.; Madramootoo, C.A. Machine Learning for Predicting Greenhouse Gas Emissions from Agricultural Soils. Sci. Total Environ.; 2020; 741, 140338.

22. Lin, G.; Huang, P.; Wang, M.; Xu, Y.; Zhang, R.; Zhu, L. An Inverse Kinematics Solution for a Series-Parallel Hybrid Banana-Harvesting Robot Based on Deep Reinforcement Learning. Agronomy; 2022; 12, 2157. [DOI: https://dx.doi.org/10.3390/agronomy12092157]

23. Bu, F.; Wang, X. A Smart Agriculture IoT System Based on Deep Reinforcement Learning. Future Gener. Comput. Syst.; 2019; 99, pp. 500-507.

24. Moghadas, D. One-Dimensional Deep Learning Inversion of Electromagnetic Induction Data Using Convolutional Neural Network. Geophys. J. Int.; 2020; 222, pp. 247-259. [DOI: https://dx.doi.org/10.1093/gji/ggaa161]

25. Xie, M.; Li, G.; Ma, W.; Liu, S.; Chang, H.; Du, J.; Ding, N. Effects of water treatment on stoichiometric ratio and homeostasis of silage corn soil in Hexi. Arid Zone Res.; 2022; 39, pp. 1312-1321.

26. Zhang, H. Uncrewed Aircraft Systems (UAS)-Based Detection of Organic Carbon with Hyper- and Multispectral Sensors in Cropland Soils. Ph.D. Thesis; Université Catholique de Louvain: Louvain-la-Neuve, Belgium, 2022.

27. Wang, Y.; Kootstra, G.; Yang, Z.; Khan, H. UAV Multispectral Remote Sensing for Agriculture: A Comparative Study of Radiometric Correction Methods Under Varying Illumination Conditions. Biosyst. Eng.; 2024; 248, pp. 240-254. [DOI: https://dx.doi.org/10.1016/j.biosystemseng.2024.11.005]

28. Stott, E.; Williams, R.D.; Hoey, T.B. Ground Control Point Distribution for Accurate Kilometre-Scale Topographic Mapping Using an RTK-GNSS Unmanned Aerial Vehicle and SfM Photogrammetry. Drones; 2020; 4, 55. [DOI: https://dx.doi.org/10.3390/drones4030055]

29. Chi, J.; Kim, J.I.; Lee, S.; Jeong, Y.; Kim, H.C.; Lee, J.; Chung, C. Geometric and Radiometric Quality Assessments of UAV-Borne Multi-Sensor Systems: Can UAVs Replace Terrestrial Surveys?. Drones; 2023; 7, 411. [DOI: https://dx.doi.org/10.3390/drones7070411]

30. IDB. Index DataBase. A Database for Remote Sensing Indices. The IDB Project, 2011–2024. Available online: https://www.indexdatabase.de/ (accessed on 8 September 2024).

31. Fernandez-Buces, N.; Siebe, C.; Cram, S.; Palacio, J.L. Mapping soil salinity using a combined spectral response index for bare soil and vegetation: A case study in the former lake Texcoco, Mexic. J. Arid Environ.; 2006; 65, pp. 644-667.

32. Datt, B.; McVicar, T.R.; Van Niel, T.G.; Jupp, D.L.; Pearlman, J.S. Preprocessing EO-1 Hyperion hyperspectral data to support the application of agricultural indexes. IEEE Trans. Geosci. Remote Sens.; 2003; 41, pp. 1246-1259. [DOI: https://dx.doi.org/10.1109/TGRS.2003.813206]

33. Roujean, J.L.; Breon, F.M. Estimating PAR absorbed by vegetation from bidirectional reflectance measurements. Remote Sens. Environ.; 1995; 51, pp. 375-384. [DOI: https://dx.doi.org/10.1016/0034-4257(94)00114-3]

34. Wang, F.M.; Huang, J.F.; Tang, Y.L.; Wang, X.Z. New vegetation index and its application in estimating leaf area index of rice. Rice Sci.; 2007; 14, pp. 195-203. [DOI: https://dx.doi.org/10.1016/S1672-6308(07)60027-4]

35. Hoffmann, H.; Jensen, R.; Thomsen, A.; Nieto, H.; Rasmussen, J.; Friborg, T. Crop water stress maps for an entire growing season from visible and thermal UAV imagery. Biogeosciences; 2016; 13, pp. 6545-6563. [DOI: https://dx.doi.org/10.5194/bg-13-6545-2016]

36. Underwood, E.; Ustin, S.; DiPietro, D. Mapping nonnative plants using hyperspectral imagery. Remote Sens. Environ.; 2003; 86, pp. 150-161. [DOI: https://dx.doi.org/10.1016/S0034-4257(03)00096-8]

37. Gitelson, A.; Merzlyak, M.N. Quantitative estimation of chlorophyll-a using reflectance spectra: Experiments with autumn chestnut and maple leaves. J. Photochem. Photobiol. B; 1994; 22, pp. 247-252. [DOI: https://dx.doi.org/10.1016/1011-1344(93)06963-4]

38. Gitelson, A.; Merzlyak, M.N. Spectral reflectance changes associated with autumn senescence of Aesculus hippocastanum L. and Acer platanoides L. leaves. Spectral features and relation to chlorophyll estimation. J. Plant Physiol.; 1994; 143, pp. 286-292. [DOI: https://dx.doi.org/10.1016/S0176-1617(11)81633-0]

39. Li, Z.; Zhou, X.; Cheng, Q.; Fei, S.; Chen, Z. A Machine-Learning Model Based on the Fusion of Spectral and Textural Features from UAV Multi-Sensors to Analyse the Total Nitrogen Content in Winter Wheat. Remote Sens.; 2023; 15, 2152. [DOI: https://dx.doi.org/10.3390/rs15082152]

40. Ni, G.; Guan, Y.; Zhang, X.; Yang, Y.; Li, Y.; Liu, X.; Rong, Z. Selection of Landsat 8 OLI Levels, Monthly Phases, and Spectral Variables on Identifying Soil Salinity: A Study in the Yellow River Delta. Appl. Sci.; 2025; 15, 2747. [DOI: https://dx.doi.org/10.3390/app15052747]

41. Da, Q.; Yan, J.; Li, G.; Guo, Z.; Li, H.; Wang, W.; Li, J.; Ma, W.; Li, X.; Cheng, K. Inversion of Soil Moisture Content in Silage Corn Root Zones Based on UAV Remote Sensing. Agriculture; 2025; 15, 331. [DOI: https://dx.doi.org/10.3390/agriculture15030331]

42. Jin, Z.; Liu, H.; Cao, H.; Li, S.; Yu, F.; Xu, T. Hyperspectral Remote Sensing Estimation of Rice Canopy LAI and LCC by UAV Coupled RTM and Machine Learning. Agriculture; 2024; 15, 11. [DOI: https://dx.doi.org/10.3390/agriculture15010011]

43. Guo, F.; Huang, Z.; Su, X.; Li, Y.; Luo, L.; Ba, Y. Soil Moisture Content Inversion Model on the Basis of Sentinel Multispectral and Radar Satellite Remote Sensing Data. J. Soil Sci.; 2024; 24, pp. 7919-7933. [DOI: https://dx.doi.org/10.1007/s42729-024-02087-z]

44. Yang, J.; Li, X.; Ma, X.; Ge, X. Potential of Hyperspectral Data Combined With Optimal Band Combination Algorithm for Estimating Soil Organic Carbon Content in Lakeside Oasis. Land Degrad. Dev.; 2024; 35, pp. 5874-5885. [DOI: https://dx.doi.org/10.1002/ldr.5339]

45. Cheng, J.K.; Feng, X.L.; Chen, L.B.; Gao, T.Y.; Mei-Jin, D.U.; Liu, Z.Y. Optimal Inversion Model for Cultivated Land Soil Salinity Based on UAV Hyperspectral Data. Ying Yong Sheng Tai Xue Bao; 2024; 35, pp. 3085-3094.

46. He, W.; Yao, Y.; Li, Y.; Yu, L.; Ni, L.; Fu, B.; Huang, J.; Li, D. Accurately Estimating and Revealing the Patterns of Leaf C:N:P Stoichiometry with Remote Sensing and Neural Network Methods in a Karst Area. Comput. Electron. Agric.; 2025; 231, 110006.

47. Liu, L.; Wei, G.; Zhou, P. Prediction and Mapping of Soil Total Nitrogen Using GF-5 Image Based on Machine Learning Optimization Modeling. Smart Agric.; 2024; 6, 61.

48. Ramírez, P.B.; Calderón, F.J.; Jastrow, J.D.; Ping, C.L.; Matamala, R. Applying NIR and MIR Spectroscopy for C and Soil Property Prediction in Northern Cold-Region Ecosystems. Which Approach Works Better?. Geoderma Reg.; 2023; 32, e00617.

49. Serrano, J.; Shahidian, S.; Paixão, L.; Marques da Silva, J.; Paniágua, L.L. Pasture Quality Assessment Through NDVI Obtained by Remote Sensing: A Validation Study in the Mediterranean Silvo-Pastoral Ecosystem. Agriculture; 2024; 14, 1350. [DOI: https://dx.doi.org/10.3390/agriculture14081350]

50. Yang, N.; Zhang, Z.; Ding, B.; Wang, T.; Zhang, J.; Liu, C.; Zhang, Q.; Zuo, X.; Chen, J.; Cui, N. et al. Evaluation of Winter-Wheat Water Stress with UAV-Based Multispectral Data and Ensemble Learning Method. Plant Soil; 2024; 491, pp. 647-668.

51. Joiner, J.; Yoshida, Y.; Anderson, M.; Holmes, T.; Hain, C.; Reichle, R.; Koster, R.; Middleton, E.; Zeng, F.W. Global Relationships Among Traditional Reflectance Vegetation Indices (NDVI and NDII), Evapotranspiration (ET), and Soil Moisture Variability on Weekly Time-scales. Remote Sens. Environ.; 2018; 219, pp. 339-352.

52. Elshikha, D.E.M.; Hunsaker, D.; Bronson, K.; Sanchez, P. Using RGB-Based Vegetation Indices for Monitoring Guayule Biomass, Moisture Content, and Rubber. Proceedings of the 2016 ASABE Annual International Meeting; Orlando, FL, USA, 17–20 July 2016.

53. Deng, B.; Yang, W.N.; Mu, N.; Zhang, C. The Research of Vegetation Water Content Based on Spectrum Analysis and Angle Slope Index. Guang Pu Xue Yu Guang Pu Fen Xi; 2016; 36, pp. 2546-2552.

54. Harris, A.; Bryant, R.G.; Baird, A.J. Detecting Near-Surface Moisture Stress in Sphagnum spp. Remote Sens. Environ.; 2005; 97, pp. 371-381.

55. Serrano, J.; Shahidian, S.; Marques da Silva, J.; Paixão, L.; Carreira, E.; Carmona-Cabezas, R.; Nogales-Bueno, J.; Rato, A.E. Evaluation of Near Infrared Spectroscopy (NIRS) and Remote Sensing (RS) for Estimating Pasture Quality in Mediterranean Montado Ecosystem. Appl. Sci.; 2020; 10, 4463. [DOI: https://dx.doi.org/10.3390/app10134463]

56. Klemas, V.; Smart, R.; Klemas, V. The Influence of Soil Salinity, Growth Form, and Leaf Moisture on the Spectral Radiance of Vegetation. Photogramm. Eng. Remote Sens.; 1983; 49, pp. 77-83.

57. Wu, Q.; Zhu, Z.; Ma, J.; Liu, S.; Chai, L.; Xu, Z. Soil Moisture Retrieval Based on Ensemble Learning Models Using Landsat8 Data in Areas of High Heterogeneity. Preprints; 2024; 2024120464. [DOI: https://dx.doi.org/10.20944/preprints202412.0464.v1]

58. Sharma, S.; Khanal, P. Predicting Above-Ground Biomass of Forest in South Carolina: Integrating Remote Sensing, Machine Learning, and Interpolation Techniques. Preprints; 2024; 2024091867. [DOI: https://dx.doi.org/10.20944/preprints202409.1867.v1]

59. Belloli, T.F.; de Arruda, D.C.; Guasselli, L.A.; Cunha, C.S.; Korb, C.C. Modeling Wetland Biomass and Aboveground Carbon: Influence of Plot Size and Data Treatment Using Remote Sensing and Random Forest. Land; 2025; 14, 616. [DOI: https://dx.doi.org/10.3390/land14030616]

60. Kim, Y.I.; Park, W.H.; Shin, Y.; Park, J.-W.; Engel, B.; Yun, Y.-J.; Jang, W.S. Applications of Machine Learning and Remote Sensing in Soil and Water Conservation. Hydrology; 2024; 11, 183. [DOI: https://dx.doi.org/10.3390/hydrology11110183]

61. Sarkar, S.; Leyton, J.M.O.; Noa-Yarasca, E.; Adhikari, K.; Hajda, C.B.; Smith, D.R. Integrating Remote Sensing and Soil Features for Enhanced Machine Learning-Based Corn Yield Prediction in the Southern US. Sensors; 2025; 25, 543. [DOI: https://dx.doi.org/10.3390/s25020543]

62. Nadeem, A.A.; Zha, Y.; Shi, L.; Ali, S.; Wang, X.; Zafar, Z.; Afzal, Z.; Tariq, M.A.U.R. Spatial Downscaling and Gap-Filling of SMAP Soil Moisture to High Resolution Using MODIS Surface Variables and Machine Learning Approaches over ShanDian River Basin, China. Remote Sens.; 2023; 15, 812. [DOI: https://dx.doi.org/10.3390/rs15030812]

63. Vinci, A.; Baiocchi, V.; Brigante, R.; Traini, C.; Farinelli, D. Use of Drones with Multispectral and Thermal Cameras to Assess the Biometric Characteristics and Water Status of Different Hazelnut Cultivars. Preprints; 2024; 2024031201. [DOI: https://dx.doi.org/10.20944/preprints202403.1201.v1]

64. Dong, H.; Dong, J.; Sun, S.; Bai, T.; Zhao, D.; Yin, Y.; Shen, X.; Wang, Y.; Zhang, Z.; Wang, Y. Crop Water Stress Detection Based on UAV Remote Sensing Systems. Agric. Water Manag.; 2024; 303, 109059.

65. Yang, S.; Li, L.; Fei, S.; Yang, M.; Tao, Z.; Meng, Y.; Xiao, Y. Wheat Yield Prediction Using Machine Learning Method Based on UAV Remote Sensing Data. Drones; 2024; 8, 284. [DOI: https://dx.doi.org/10.3390/drones8070284]

66. Liu, J.; Xiang, J.; Jin, Y.; Liu, R.; Yan, J.; Wang, L. Boost Precision Agriculture with Unmanned Aerial Vehicle Remote Sensing and Edge Intelligence: A Survey. Remote Sens.; 2021; 13, 4387. [DOI: https://dx.doi.org/10.3390/rs13214387]