A robust regional soil moisture estimation method based on spatiotemporal two-stage data-driven models

Soil moisture (SM) is a key variable in the hydrological cycle and agricultural production. Departing from traditional physics-based models, this study accounted for distinct horizontal and vertical drivers of soil moisture and proposed a spatiotemporal two-stage LSTM (LSTM-ST) for reconstructing regional 3D soil moisture dynamics. The first stage (LSTM-S) estimates SM at any horizontal location by integrating dynamic and static features across monitoring stations. The second stage (LSTM-T) captures vertical lag effects of rainfall on SM at different depths using multi-depth time series data. The proposed LSTM-ST was evaluated using datasets from 115 sites across the USA, with Backpropagation (BP) neural networks serving as a benchmark. The results demonstrated that LSTM-ST can achieve high estimation accuracy in both horizontal and vertical dimensions. Due to its high sensitivity to meteorological conditions, surface SM estimation accuracy exhibited a non-monotonic trend with increasing training site density, whereas deep-layer SM prediction accuracy, governed by time-lag effects, first increased and then decreased. Shorter sequences were recommended when data volume is sufficient. Soil classification generally improved model accuracy, but insufficient post-classification sample size risked inducing model instability and subsequent accuracy degradation. Shorter input sequences were preferred when sufficient training data were available. Soil classification generally improved accuracy, although small post-classification samples could introduce instability. Overall, the two-stage structure promotes temporal continuity and physical consistency, enabling strong generalization and interpretability, and supports applications in hydrological analysis, precision irrigation, and extreme weather response.