Research on Soil Moisture Prediction Based on LSTM-Transformer Model

Soil moisture is one of the basic climate variables of the global climate observation system, and the prediction of soil moisture is of great significance for agricultural yield assessment, flood and drought prediction, and soil and water conservation. Aiming at the complexity of soil moisture influencing factors and their time-varying time series characteristics, we propose a Transformer model that introduces LSTM, which uses the sequential modeling capability of LSTM to extract contextual information for each data, and plays the role of position coding in the LSTM-Transformer model, and the multi-head attention mechanism in the model can highlight important features by weighting, so as to effectively process time series data. Taking soil moisture, soil evaporation, vegetation index, runoff and climate data at different depths of Xilin Gol grassland in Inner Mongolia from 2012 to 2022 as input variables, soil moisture at different depths from 2022 to 2023 was predicted, and the model prediction performance was compared with the traditional long short-term memory neural network (LSTM) and bidirectional long short-term memory neural network (BiLSTM) through the three statistical indicators of MAE, MAPE and RMSE. The LSTM-Transformer model has better performance for prediction of soil moisture at different depths. The prediction of soil moisture has great guiding significance for timely grasping grassland soil moisture and adopting proactive agricultural production.