A novel diversity-aware sampling method for global soil moisture prediction

Abstract

Predicting global soil moisture (SM) is crucial for drought forecasting, agricultural management, and climate modeling. However, traditional deep learning (DL) methods often struggle with imbalanced sample distributions and limited spatial representation, which restrict their ability to accurately model SM patterns across diverse regions and time. Furthermore, variations in sample characteristics influenced by spatial proximity pose additional challenges in creating balanced and representative training datasets. To address these challenges, we propose a Diversity-Aware Sampling (DAS) strategy to enhance spatial representativeness and temporal diversity in training data. DAS enhances traditional sampling by grouping samples through clustering and categorizing each cluster into high, medium, and low uncertainty levels. This approach ensures each batch contains a balanced mix of samples across multiple grid points, improving coverage and representativeness. Applied to an LSTM-based model for 1- to 3-day global SM predictions, DAS achieved notable performance gains, increasing R2 by up to 8.39% and KGE by up to 6.38%, demonstrating improved accuracy and stability. Ground-based evaluations using China Meteorological Administration (CMA) station data for 5-day drought forecasting further validated DAS’s superiority over traditional methods. By improving the spatial and temporal representativeness of training samples, DAS enhances the generalization of deep learning models in geoscience applications. This robust framework offers significant advancements in global SM prediction and drought monitoring.