Coupling time-series decomposition with causal analysis enhances soil moisture prediction in humid climates

Soil moisture is a key variable linking atmospheric processes and crop growth, and its accurate prediction is essential for agricultural management, water resource allocation, and ecological security. However, traditional prediction methods often fail to capture nonlinear dynamics and complex dependencies among multi-source drivers. Focusing on the humid Poyang Lake Basin, this study integrates in-situ soil moisture observations with Soil Moisture Active and Passive (SMAP) satellite products to identify effective stations with strong ground–satellite consistency. The SMAP soil moisture time series was decomposed using Seasonal–Trend decomposition based on Loess, and causal structure analyses were conducted between meteorological variables and the trend, seasonal, and residual components. An improved prediction framework was developed by coupling time-series decomposition and causal modeling within a Long Short-Term Memory network (LSTM). The spatiotemporal performance of 7-day soil moisture forecasts was evaluated through different ablation experiments. Results show that the hybrid deep learning model consistently outperforms the benchmark models, with the median root mean square error of 4.94% with 95% confidence interval of 4.86%–5.24% and the Kling–Gupta efficiency of 0.71 (0.69–0.75), respectively. Spatially, higher accuracy is observed in the central plains and northern lakeside regions, while conventional LSTM performs relatively better in mountainous areas. The performance gains mainly stem from explicit multi-scale temporal feature extraction and improved representation of meteorological driving mechanisms. This study provides a robust framework for basin-scale soil moisture prediction and supports drought monitoring and water management in humid and semi-humid regions.