Abstract
<title>Abstract</title> <p>To enable rapid and effective monitoring of the spatiotemporal dynamics of soil moisture (SM) across different growth stages and soil depths in spring maize fields in Northwest China, this study integrates optical vegetation indices (VIs) with radar backscatter features (RBFs) within a machine learning framework. The contributions of different features and their combinations to soil moisture prediction across various soil layers were analyzed, and the estimation potential of three models, XGBoost, SVR, and KNN, during key growth stages of spring maize was investigated. The results indicated that: (1) VIs and RBFs exhibited complementarity in soil moisture retrieval. Surface soil moisture was more sensitive to radar backscatter, whereas deeper soil moisture showed a stronger correlation with vegetation indices. As the growth stage progressed, the correlation between surface moisture and radar features gradually decreased, while the correlation between deeper moisture and radar features first increased and then decreased. (2) The XGBoost model consistently outperformed SVR and KNN across all growth stages, soil depths, and input feature combinations, achieving validation R² values ranging from 0.525 to 0.825, with significantly superior generalization capability. (3) The combined VI + RBF feature set outperformed single-feature inputs at all growth stages and depths, substantially improving soil moisture estimation accuracy. (4) SHAP-based global interpretation revealed that VV polarization served as the core driving feature for surface soil moisture estimation, whereas vegetation indices were more important for estimating deeper soil moisture. (5) Soil moisture exhibited pronounced spatiotemporal heterogeneity. At each growth stage, soil moisture increased with soil depth, with substantial surface variability and stable distribution at deeper layers. Temporally, soil moisture followed a unimodal trend across the growing season, first increasing and then decreasing, peaking at the tasseling stage and reaching its minimum at the jointing stage. This approach enables effective spatiotemporal dynamic monitoring of soil moisture in spring maize fields.</p>