Prediction of Farmland Soil Moisture at Different Depths in the Semiarid Region of Northern China Using the LSTM and DT Models

Abstract

Accurate prediction of farmland soil moisture is crucial for determining crop water requirements and establishing effective irrigation standards. However, the high costs and potential disruption of soil structure make direct measurement of soil moisture across various depths challenging. In this study, the long short-term memory (LSTM) and decision tree (DT) models were proposed to predict soil moisture at 3-, 5-, 10- and 20-cm soil depths based on soil temperature data, and the accuracies of the two models in predicting soil moisture at different depths were evaluated at half-hour and daily scales. The results revealed that the accuracies of the LSTM and DT models in predicting soil moisture at different depths at the half-hour scale were greater than those at the daily scale. The accuracy of the LSTM model was better than that of the DT model at different depths. Both models performed best at the 20-cm soil depth, followed by the 10-, 5- and 3-cm soil depths, with R² values ranging from 0.90–0.95, 0.81–0.95, 0.84–0.95 and 0.86–0.95, respectively. Therefore, the LSTM model is recommended for the prediction of soil moisture at different soil depths, providing valuable references for farmland management, irrigation decision-making and the formulation of drought prevention and mitigation measures.