Soil moisture prediction research in the Beijing-Tianjin-Hebei region based on particle swarm optimized extreme learning machine

Soil moisture plays a vital role in ecological dynamics, the hydrological cycle, and climate monitoring. Traditional prediction methods include physical models and data-driven approaches, with the latter gaining popularity due to its ability to capture complex data patterns. This study utilizes the Extreme Learning Machine (ELM) model to predict surface soil moisture (0–5 cm) by incorporating meteorological, topographic, and soil texture data. To improve model stability and accuracy, we integrate the Particle Swarm Optimization (PSO) algorithm for global parameter optimization. Results indicate that incorporating static features, such as topography and soil texture, significantly enhances prediction accuracy. The particle swarm optimization–extreme learning machine (PSO-ELM) model achieves a 14.6% increase in the coefficient of determination (R²), with reductions of 27.5% in root-mean-square error (RMSE) and 23.8% in mean absolute error (MAE). Validation with ground-based observations demonstrates superior performance. The PSO-ELM model provides high-resolution, accurate soil moisture predictions, supporting agricultural management and land surface process studies.