Yu et al. (2025) Estimating Winter Wheat Leaf Water Content by Combining UAV Spectral and Texture Features with Stacking Ensemble Learning

Identification

• Journal: Agronomy
• Year: 2025
• Date: 2025-11-13
• Authors: Xingjiao Yu, Long Qian, Kainan Chen, Sumeng Ye, Qi Yin, Lin Shao, Danjie Ran, Wenè Wang, Baozhong Zhang, Xiaotao Hu
• DOI: 10.3390/agronomy15112610

Research Groups

• Key Laboratory of Agricultural Soil and Water Engineering in Arid and Semiarid Areas, Ministry of Education, Northwest AF University, Yangling, China
• College of Water Resources and Architectural Engineering, Northwest AF University, Yangling, China
• State Key Laboratory of Simulation and Regulation of Water Cycle in River Basin, China Institute of Water Resources and Hydropower Research, Beijing, China
• National Center for Efficient Irrigation Engineering and Technology Research-Beijing, Beijing, China

Short Summary

This study developed a stacking ensemble learning model integrating UAV multispectral and texture features to accurately estimate winter wheat leaf water content (LWC). The model achieved significantly improved estimation accuracy (R² = 0.865, rRMSE = 16.3%) compared to single-feature or single-model approaches, demonstrating the effectiveness of multi-source feature fusion for precision agricultural water monitoring.

Objective

• To develop an inversion model for winter wheat leaf water content (LWC) based on a stacking ensemble learning framework, integrating multispectral and texture features to improve estimation accuracy.
• To construct a fused spectral–texture feature dataset from UAV multispectral imagery.
• To evaluate the performance of the stacking algorithm-based estimation model.
• To generate spatial distribution maps of winter wheat LWC for precise water management.

Study Configuration

• Spatial Scale: Field-scale (Wugong County, Shaanxi Province, China); 90 sampling points systematically arranged in 30 m × 30 m quadrats; UAV imagery with centimeter-level spatial accuracy (flight altitude 100 m; 3 × 3 window for texture feature extraction).
• Temporal Scale: March to May 2024, covering the regreening and jointing stages of winter wheat. Data collection occurred during midday (10:00–14:00) to minimize illumination variation.

Methodology and Data

• Models used:
  • Base learners: Support Vector Machine (SVM), Random Forest (RF), Extreme Gradient Boosting (XGBoost), K-Nearest Neighbors (KNN), Partial Least Squares Regression (PLSR).
  • Meta-learner: Linear Regression (LR).
  • Ensemble framework: Stacking ensemble learning.
• Data sources:
  • UAV multispectral imagery: Collected using a Mavic 3 Multispectral drone, capturing green (560 ± 16 nm), red (650 ± 16 nm), red edge (730 ± 10 nm), and near-infrared (860 ± 26 nm) bands.
  • Ground truth data: Destructive sampling of winter wheat plants (90 points) to measure fresh weight (mf) and dry weight (md) using the oven-drying method (ISO 11465:1993) for leaf water content (LWC) calculation.
  • Derived features: 17 Vegetation Indices (VIs) (e.g., NDVI, NDWI, SAVI, TVI) and 32 Gray Level Co-occurrence Matrix (GLCM) texture features (TFs) (e.g., Mean, Variance, Homogeneity, Contrast, Dissimilarity, Entropy, Angular Second Moment, Correlation) extracted from the four spectral bands.

Main Results

• Models using only multispectral features yielded R² values of 0.526–0.718 and rRMSE of 22.795–29.536%.
• Models relying solely on texture features performed worse, with R² values of 0.273–0.425 and rRMSE of 34.7–36.6%.
• Combining multispectral and texture features notably improved accuracy, achieving R² values of 0.748–0.815 and rRMSE of 18.5–21.6%.
• The stacking ensemble learning model, when applied to the fused features, achieved the highest precision with R² = 0.865 and rRMSE = 16.3%.
• Feature importance analysis revealed that near-infrared (NIR) related spectral indices (e.g., SAVI, RDVI, TVI) and texture features derived from red-edge and NIR bands (e.g., contrast, homogeneity, variance) provided the most critical information for LWC estimation.
• Spatial distribution maps of LWC derived from the stacking model effectively revealed field-scale moisture differences and spatial heterogeneity, showing higher variability during the regreening stage compared to the jointing stage.

Contributions

• Developed a novel and highly accurate inversion model for winter wheat LWC by integrating UAV multispectral and texture features within a stacking ensemble learning framework.
• Demonstrated the complementary advantages of combining spectral (capturing biochemical responses) and spatial texture (reflecting canopy structure and heterogeneity) information for LWC estimation, addressing a critical research gap.
• Showcased the superior predictive performance and robustness of stacking ensemble learning over single machine learning models for crop water status monitoring in complex agricultural environments.
• Provided a reliable and scalable approach for precision agricultural water monitoring, offering a scientific basis for optimizing irrigation timing and resource use.

Funding

• National Natural Science Foundation of China (U2243235)
• National Key Research and Development Program of China (2022YFD1900402-01)

Citation

@article{Yu2025Estimating,
  author = {Yu, Xingjiao and Qian, Long and Chen, Kainan and Ye, Sumeng and Yin, Qi and Shao, Lin and Ran, Danjie and Wang, Wenè and Zhang, Baozhong and Hu, Xiaotao},
  title = {Estimating Winter Wheat Leaf Water Content by Combining UAV Spectral and Texture Features with Stacking Ensemble Learning},
  journal = {Agronomy},
  year = {2025},
  doi = {10.3390/agronomy15112610},
  url = {https://doi.org/10.3390/agronomy15112610}
}