Depth-dependent predictive information in monitored soil moisture profiles: Contrasting roles of persistence, forcing, and HYDRUS-derived states

Process-based simulations, recent observations, and external forcing may contain overlapping information for soil-water estimation, making the incremental value of process-model states difficult to establish. We evaluated daily soil-water-state estimation and HYDRUS-1D post-processing at ten depths (10–100 cm) in one monitored tea-plantation profile and one grassland profile in northern China. Nine model formulations were assessed using chronological holdouts. Persistence, defined as the previous-day observed soil water content (SWC), served as an explicit monitoring-based benchmark. The primary comparison used equal-target Light Gradient Boosting Machine (LightGBM) models with identical non-HYDRUS inputs, differing only in the inclusion of five HYDRUS-derived states. All depths used the same candidate lags and windows, and uncertainty was evaluated using 1000 paired 7-day block-bootstrap replicates. Shapley-based model (SHAP) attribution was aggregated across all effective features and all test days. In the tea plantation, Persistence had lower root-mean-square error (RMSE) than both Direct LightGBM models at all ten depths. Adding HYDRUS produced no bootstrap-supported improvement at any depth, while HYDRUS-family attribution ranged from 1.4% to 7.8%. In the grassland, the prespecified 90-day protocol supported incremental HYDRUS improvements at 50–80 cm, with the strongest cross-window directional consistency at 50–60 cm. At 40 cm, HYDRUS attribution was approximately 29.3%, but the 95% bootstrap interval for ΔRMSE crossed zero; at 90–100 cm, Persistence was markedly more accurate. HYDRUS-derived information therefore showed depth- and context-conditional incremental value. Persistence performance, incremental predictive performance, and fitted-model attribution represent distinct forms of evidence and should not be interpreted interchangeably or causally.