Field-scale soil moisture over Hungary under non-stationary drought transfer: a unified account across surface, region, and depth

Fehér Zsolt Zoltán 

Three questions decide whether a satellite-driven soil-moisture estimator is fit for operational drought monitoring over a heterogeneous country: how accurately can the surface layer be recovered, whether the controlling processes differ across the landscape, and how far the surface signal reaches into the profile that actually matters for plants and recharge. The program began with a national surface-moisture benchmark over the 133-station Hungarian OVF Aszálymonitoring network and grew from there into a regional and a depth-resolved study; all three fuse Sentinel-1 C-band SAR, harmonized Landsat/Sentinel-2 (HLS) optical data, and a rich ancillary covariate set, evaluated under a single strict protocol that withholds the 2025–2026 drought years for inter-annual transfer.
The three component studies map three orthogonal axes. On the horizontal accuracy axis, a tuned gradient-boosted tree ensemble (XGBoost) was the most accurate model for 10-cm surface moisture (RMSE 0.0478 m3 m−3, R2 0.735), beating nine deep-learning architectures and the most seed-stable of all. On the spatial-structure axis, the single national model was the best estimator in every one of five contrasting regions — region-specific retraining never lowered error (pooled regionalization gain −0.0019 m3 m−3) — yet each region carried a distinct dominant control, with selective state-space (Mamba) rewarded by local training only in the recharge-memory Nyírség and a recurrent gate (LSTM) only in the seasonal-memory Tikevir. On the vertical-observability axis, the surfacecoupled signal decayed monotonically with depth to an empirical decoupling crossover near 30 cm, governed by a storage-deficit-gated wetting front (≈10 cm penetration per 15–20 mm of rain; ≥45 cm reached in only ~20% of events) rather than static texture (R2 0.04), and the trees-win result extended down the full profile against a fairly tuned modern deep benchmark. Two laws unify the program: gradient-boosted trees beat deep learning under non-stationary drought transfer because the task is extrapolation, not interpolation; and a single broad-network model is the deployable choice (data synergy) even though the underlying processes differ by region and decay with depth. Remote sensing was redundant given the covariate-rich setting. The three axes together map the full field-scale soil-moisture problem.