Machine learning-based seasonal SMAP soil moisture retrieval integrating MODIS drought indices: A case study of the Wujiang River Basin.

Soil moisture (SM) is a critical regulator of energy and water exchange between the land and atmosphere, yet its accurate retrieval in complex Karst terrains remains challenging due to extreme surface heterogeneity and intricate hydro-thermal coupling. Traditional unified modeling approaches often struggle to capture the seasonally varying, non-linear relationships between remote sensing signals and moisture dynamics in fragmented landscapes. To address these limitations, this study developed a seasonally decoupled machine learning framework, utilizing long-term data from 2019 to 2024, to enhance the spatial representativeness of 9 km SMAP products in the Wujiang River Basin. By integrating 14 MODIS-derived drought indicators with static topographic factors, we constructed differentiated seasonal input sets to account for the "signal decoupling" potentially caused by intense precipitation pulses and phenological shifts. Five algorithms, including RBFNN, SVM, RF, XGBoost, and CatBoost, were systematically evaluated under the Optuna optimization framework. Quantitative results suggest that the CatBoost-based decoupled model achieved an improved balance of accuracy and robustness, with [Formula: see text] values reaching 0.537 and 0.572 in spring and autumn, respectively, showing certain advantages over traditional baseline models. Importantly, SHAP-based attribution analysis identifies statistical patterns consistent with a transition in hydrological behaviors, indicating that the model's decision logic shifts from a 'topography-driven' importance in winter to a 'hydro-thermal-driven' dominance in summer. This seasonally adaptive approach contributes to mitigating systematic biases in satellite products and provides a potential methodological reference for drought monitoring and water resource management in complex mountainous environments.