Machine learning modeling of spatially distributed soil moisture time series for a hillside at Sulmachun watershed, South Korea

Study region The study focuses on a steep, 4000 m² headwater hillslope within South Korea's Sulmachun Catchment. This monsoon-influenced region receives 1636 mm of annual rainfall (70% during summer) and features sandy loam soils under a mixed forest canopy. Study focus The research evaluates the predictability of spatially distributed soil moisture time series using MLP, 1D-CNN, and LSTM machine learning models. By analyzing soil moisture measurements across vertical and transactional profiles, the study aims to determine the optimal temporal lag (memory effect) for accurate predictions and to map underlying hydrological processes based on model performance. New hydrological insights for the region The LSTM model outperformed others by effectively capturing complex temporal dependencies. The modeling revealed pronounced spatial heterogeneity in hydrological memory across the hillslope. A short term 6 h antecedent lag was optimal for midslope regions, reflecting rapid vertical infiltration and post rainfall storage release. Conversely, slope boundaries required longer lags of 24–48 h. Upslope areas exhibited extended lags due to delayed infiltration and macropore flow, while downslope areas reflected prolonged saturation and direct interactions with the riparian water table. Ultimately, these data driven AI structures successfully capture complex, localized subsurface flow mechanisms and threshold behaviors on mountain hillslopes.