Diagnosing the Applicability Boundary of Water Cloud Model-Based Soil Moisture Retrieval Under Cross-Site Generalization

The Water Cloud Model (WCM) is the most widely used vegetation-scattering correction framework for microwave soil-moisture (SM) remote sensing, yet its cross-site generalization ability remains poorly understood. This study compares WCM-based random forest models across four validation networks spanning a landscape-heterogeneity gradient (REMEDHUS, SMOSMANIA, Little Washita, and OzNet; station-mean coefficient of variation (CV) of 0.18–0.67), under both within-station random-split and leave-one-station-out (LOSO) validation. The WCM-decoupled feature yields no significant accuracy gain under random validation (Wilcoxon p ≥ 0.168), and all models fall below a monthly climatology baseline under LOSO, consistent with a generalization failure that is more plausibly associated with station-level distributional shift than with vegetation-correction error. LOSO R2 decreases monotonically with landscape heterogeneity across the four networks, but at the subnetwork level it is robust only at the most heterogeneous network (REMEDHUS); in homogeneous landscapes, temporal transferability appears more important. A data-leakage artifact is identified and corrected in anomaly-based modeling, substantially reducing apparent gains at two of the four sites. A target-station-mean correction calibrated from only the earliest 10% of a new station’s observations lifts R2 above the climatology baseline at the three most heterogeneous networks, suggesting that much of the failure is a correctable station-baseline shift, although other factors cannot be excluded. These findings provide supporting evidence on the applicability boundary of WCM–machine-learning frameworks for network-scale soil moisture retrieval, based on the four networks examined.