Harnessing multimodal UAV data and ensemble learning to improve estimation accuracy of winter wheat water content

Rapid and accurate monitoring of plant water content (PWC) is crucial for characterizing crop water status and supporting water-efficient crop production. Previous researches on PWC estimation primarily relied on single-modality inputs or conventional models, limiting estimation accuracy. Therefore, in this study, multispectral and thermal infrared imagery of winter wheat was acquired using a UAV equipped with corresponding sensors. Spectral indices, multispectral texture indices, temperature indices, and thermal infrared texture indices were extracted as multimodal data. Machine learning techniques, including ensemble learning algorithms, were employed to develop a high-accuracy model for monitoring crop PWC. We first evaluated the potential of multimodal data for estimating PWC of winter wheat, then compared the estimation accuracy across different models, and finally quantified the relative contribution of each data modality using SHapley Additive exPlanations (SHAP). The results demonstrated that: (i) multimodal data fusion significantly enhanced the accuracy of PWC estimation. Compared to single-modality inputs, the R 2 values increased from 0.466–0.796 to 0.781–0.831, while RMSE and MAE decreased from 0.019–0.031 g·g −1 and 0.015–0.024 g·g −1 to 0.017–0.020 g·g −1 and 0.013–0.016 g·g −1 , respectively; (ii) ensemble models exhibited greater stability, with the Averaging model achieving the best performance under multimodal conditions (R 2 = 0.831, RMSE = 0.017 g·g −1 , MAE = 0.014 g·g −1 ); (iii) among the multimodal data, the multispectral texture indices contributed most significantly to PWC estimation accuracy, with a contribution rate ranging from 24.70% to 78.13%. Overall, the ensemble learning model leveraging UAV-based multimodal data would provide a reliable and robust method for estimating winter wheat PWC, supporting field-scale monitoring of crop water status.