Leaf water content-guided multi-task learning for multi-depth soil moisture estimation using multisource UAV remote sensing data

Accurate estimation of soil moisture (SM) at different depths is essential for orchard water management, yet the contribution of canopy physiological information to multi-depth SM retrieval remains insufficiently understood. In this study, we developed a multi-task learning (MTL) model that incorporates leaf water content (LWC) as an auxiliary physiological task to improve SM retrieval across varying depths (5, 10, 20 and 40 cm). LWC was significantly correlated with SM at 10, 20, and 40 cm, but not at 5 cm, suggesting that LWC may serve as a auxiliary task for deeper SM estimation. Evaluating Random Forest (RF), single-task learning (STL), and the proposed MTL models revealed distinct depth-wise performance differences. RF performed best at 5 cm (R 2 =0.685), whereas MTL achieved the highest testing accuracy at 10, 20 and 40 cm, with R 2 of 0.691, 0.695, and 0.554, respectively. Overall estimation accuracy peaked at 20 cm, while the largest relative improvements of MTL over RF were observed at 10 and 40 cm (R 2 increases of 0.181 and 0.234). Finally, SHapley Additive exPlanations (SHAP) analysis further showed that incorporating LWC as an auxiliary task was associated with a shift in feature contribution patterns, with increased importance assigned to vegetation condition, pigment-related, and thermal stress indicators. These results indicate that incorporating LWC as an auxiliary task can improve subsurface SM estimation and enhance model interpretability for precision orchard water management.