Earth observation and machine-learning–based mapping of 0–1 m soil moisture at 10-cm intervals in a permafrost-affected basin

Depth-resolved soil moisture is hard to obtain at basin scale because satellites sense only the upper few centimeters and field profiles are sparse. We use Earth-observation (EO) predictors with limited profile data to map volumetric water content (VWC) at 10-cm intervals from 0–1 m in a permafrost-affected basin. The pipeline benchmarks tree ensembles and then applies Shapley Additive Explanations (SHAP)-guided selection to build a compact 12-predictor Extra Trees model, followed by block-wise mapping on 30-m grids. On an external 10% hold-out, the model achieves R 2 = 0.85 and RMSE = 0.08 L/L, with > 80% of residuals within ± 0.10 L/L. Skill is stable through most layers but degrades at 90–100 cm, reflecting limited test support and weaker surface constraint at depth. SHAP attribution identifies permafrost presence, wetland extent, vegetation greenness, and terrain metrics as leading controls, with surface influences weaken below 50 cm. Residual diagnostics show no pronounced distance-to-river shift in residual centroids and no significant spatial clustering in site-mean residuals (Moran’s I). The resulting GIS-ready GeoTIFFs provide depth-layered, climatology-consistent thaw-season VWC patterns that complement coarse satellite soil-moisture products by resolving subsurface heterogeneity.