Evaluation of winter-wheat water stress with UAV-based multispectral data and ensemble learning method.

Aims: Combining multiple features of UAV-based multispectral images with the stacking ensemble model, to improve the feasibility and accuracy of evaluating water stress in winter wheat. Methods: UAV-based multispectral images of winter wheat with different moisture treatments were acquired, from which the features such as spectrum, texture, and color moments were extracted. The soil moisture content (SMC) as well as fuel moisture content (FMC), plant moisture content (PMC), and above-ground biomass (AGB) were collected for charging the degree of water stress. The basic models were used to build ensemble models such as stacking and weighted stacking (WE-stacking), and we estimated SMC, FMC, PMC and AGB combined with multiple features. The performance of these models was evaluated. Results: The more severe the water stress, the lower values of SMC, FMC, PMC and AGB were obtained with estimation models. The performance of estimation models based on multi-feature fusion outperformed single feature in the evaluation of winter-wheat water stress. In the estimation of SMC, both stacking and WE-stacking models performed better than the basic models. Compared to the stacking model, the WE-stacking model had higher accuracy, with R2 increased between 1.98% and 3.62% at different soil depths. The WE-stacking model with multi-feature fusion still had sufficient stability and high accuracy in FMC, PMC and AGB estimation, with R2 of 0.866, 0.881 and 0.884, respectively. Conclusions: The multi-feature fusion of UAV multispectral images combined with WE-stacking model has great application potential and provides technical support in evaluating crop water stress. [ABSTRACT FROM AUTHOR]