Global-regional nested forecasting of soil moisture

Abstract

Soil moisture is a key component of the Earth system and is important for ecosystem functioning and water resources. However, existing approaches still face limitations in soil moisture forecast skill and spatial resolution. Here, we develop an Artificial Soil Moisture Forecasting Model (ASM) with a global–regional nested framework that links global low-resolution prediction with regional high-resolution forecasting. ASM consistently outperforms representative deep learning models across forecast lead times, with ablation experiments confirming the contributions of its major architectural components. Compared with ECMWF, ASM more closely reproduces ERA5 soil moisture fields and preserves greater spatial heterogeneity at 1° resolution. At a 14 day lead time, ASM achieves an ACC of 0.612, demonstrating reliable early-subseasonal forecast skill. At the regional scale, ASM provides 0.1° soil moisture forecasts for Henan Province, China, and Southern Africa, while improving extreme drought detection relative to ECMWF-driven forecasts. Attribution analysis shows that antecedent soil moisture is the dominant predictor, accounting for 61.2% of the total attribution and highlighting the importance of soil moisture memory. Soil-moisture-only autoregressive experiments further highlight that external atmospheric forcing remains essential for maintaining forecast skill. Overall, ASM provides an scalable and interpretable framework for synoptic-to-early-subseasonal soil moisture forecasting.