Research on soil moisture prediction model based on multihead LSTM：a case study in Tianjun county in the Qinghai Lake basin

Soil moisture plays a significant role in regional water and energy cycles, and dynamic information is crucial for research on water resources management and agricultural production． Site 27 dataset with no missing values in Tianjun dense soil moisture and Freeze–Thaw monitoring network in the Qinghai Lake basin were applied to a multihead long short-term memory network (multihead LSTM)． Three input soil parameters for the network were soil moisture, soil conductivity and soil temperature, with sliding coefficients of variation． Deep learning prediction models for soil moisture were established at soil depth of 5, 10 and 30 cm to predict soil moisture after 1, 7 and 30 days． Coefficient of determination (R²) of the models averaged 0.90 under varied soil depth, with root mean square error (RMSE) at 0.031 and mean absolute percentage error (MAPE) at 15.33% in average． This work therefore performed high-precision time series prediction of multi-layer soil moisture on varied time scales．