An approach for multi-depth soil moisture prediction in alfalfa based on a dual-branch combined model

Rainfall or irrigation induces substantial fluctuations in soil moisture at various depth. Improving the accuracy of multi-depth soil moisture prediction during these events is crucial for precision irrigation. This study proposes a dual-branch combined model for multi-depth soil moisture prediction in alfalfa (ALFSMP-DBCM). The model employs fully connected layers in the left branch to extract rainfall and irrigation features, while the right branch uses convolutional residual networks to model soil moisture relationships. The fusion of these branches enables effective multi-depth soil moisture prediction for alfalfa. Field experiments were designed and conducted in the Ningxia Irrigation Area of the Yellow River (NIR). A comprehensive dataset, comprising 19,763 data points on alfalfa growth environment in the different precipitation years (2017, 2018, and 2022), was established and utilized as model training data. Three classical deep learning models were employed for comparison. Results demonstrated that the ALFSMP-DBCM model effectively predicted multi-depth soil moisture during all alfalfa growth stages. The R2 of the model within the range of 0.911 to 0.992, with average MAE, MSE, and RMSE within the range of 0.29% to 0.58%, 0.22% to 0.56%, and 0.47% to 0.68%, respectively. Compared to the ANN, LSTM, and BiLSTM models, the ALFSMP-DBCM model improved the prediction accuracy of soil moisture at multi-depth by 7.19%, 11.90%, and 10.32%, respectively. The model exhibited robust performance under instantaneous water replenishment conditions and stability in predicting multi-depth soil moisture with different delay days. These findings provide a valuable reference for precision irrigation regulation and field management of alfalfa in arid and semi-arid regions.