UAV multisource remote sensing inversion of soil moisture in a farming pastoral ecotone micro landform areas

Soil moisture is a key variable controlling vegetation growth, evapotranspiration, water redistribution, and agricultural productivity in arid and semiarid regions. Accurate retrieval of soil moisture is particularly difficult in farming-pastoral ecotones, where cropland, woodland, grassland, shrubland, and sandy land are spatially interlaced and where micro landforms strongly influence infiltration, runoff redistribution, and water retention. This study aimed to retrieve multilayer soil moisture in a typical farming pastoral ecotone of the southeastern Horqin Sandy Land by integrating unmanned aerial vehicle multisource remote sensing and machine learning. Field observations were conducted after a short rainfall event in late August 2024. Soil volumetric water content was measured at 26 plots on four dates, and synchronous UAV data were acquired from RGB, multispectral, thermal infrared, and LiDAR sensors. Soil moisture was measured for four cumulative soil layers, namely 0-3.8, 0-7.5, 0-12, and 0-20 cm. A total of 70 candidate variables were constructed from RGB color indices, multispectral vegetation indices, texture features, temperature-related indicators, vegetation temperature dryness information, and terrain attributes. Pearson correlation coefficient analysis, Boruta feature selection, and best subset selection were used to derive feature subsets, and four machine learning models, including random forest regression, gradient boosting decision tree, CatBoost, and back propagation neural network, were compared for each soil layer. Model performance was evaluated using the coefficient of determination, root mean square error, and mean absolute error. The optimal model was then applied to map the spatial distribution of soil moisture after rainfall, and SHAP analysis was used to identify the dominant variables affecting model output. The results showed that CatBoost combined with the Boruta selected feature subset achieved the best performance for all soil layers. For the 0-3.8, 0-7.5, 0-12, and 0-20 cm layers, the highest R² values were 0.89, 0.88, 0.84, and 0.70, respectively, with corresponding RMSE values of 1.04%, 0.86%, 1.35%, and 1.62%. Model accuracy generally decreased with increasing soil depth, indicating that deeper soil moisture was less directly coupled with surface spectral and thermal responses. Spatial retrieval results reproduced the rapid increase and gradual decline of soil moisture after rainfall and showed clear differences among different land cover types. SHAP analysis indicated that DEM and TVDI were the dominant variables, suggesting that micro topography- and drought-related information jointly controlled soil moisture distribution. These results demonstrate that UAV multisource remote sensing combined with Boruta and CatBoost can effectively retrieve multilayer soil moisture in heterogeneous dryland agricultural landscapes and provide technical support for refined soil moisture monitoring and agricultural water management.