Machine learning with feature selection reveals key drivers of multi-depth soil moisture content

Abstract

Soil moisture content (SMC) plays a vital role in agricultural productivity, water resource management, and ecosystem sustainability in semi-arid regions. Despite this importance, most existing machine learning models mainly rely on remote sensing data to predict the soil moisture variation in the surface soil; however, they are constrained by redundant input features and limited interpretability. To address these shortcomings, this study combines the Random Forest (RF) algorithm, Convolutional Neural Networks (CNN), and the Transformer framework to develop a hybrid RF-CNN-Transformer model. Specifically, the RF algorithm, CNN, and Transformer framework are respectively used for selecting influential features, extracting spatial patterns, and capturing long-term temporal dependencies. Applied to the Mu Us Sandy Land using data from six soil depths (5, 10, 20, 40, 70, and 87 cm), the model demonstrated high prediction accuracy and training efficiency across all layers compared to baseline models, with R² values ranging from 0.8586 to 0.984 (mean R² = 0.9507). Interpretability analysis revealed a shift in the controlling mechanisms of soil moisture: shallow-layer SMC is jointly influenced by meteorological conditions and groundwater level, whereas groundwater becomes the dominant factor in deeper layers. Notably, due to the extremely dry climate, precipitation has a relatively minor impact on soil moisture dynamics across all depths. Overall, the proposed RF-CNN-Transformer model enhances both the predictive capability and interpretability of soil moisture variation, supporting precision irrigation and water resource optimization in agriculture, especially in arid and semi-arid regions.