A multihead LSTM technique for prognostic prediction of soil moisture

Abstract

Prognostic prediction of soil moisture is a critical step in various fields such as geotechnical engineering, agriculture, geology, hydrology, and climatology. For example, in agricultural applications, soil moisture prediction is needed one-day and one-week ahead of time to optimize crop planting quality and set irrigation schedule, respectively. In soil and environmental management applications, soil moisture prediction is usually needed one-month ahead of time. Therefore, a capability to accurately forecast soil moisture is of paramount importance. For this purpose, deep learning methods, especially the long short-term memory network (LSTM) method, have been used extensively. It is shown that such models can successfully forecast the soil moisture for a short time frame in future, but their accuracy sharply decreases for long time frames. To resolve this issue, in this study, we present a multihead LSTM model that learns a number of hypotheses and aggregates them for prognostic prediction tasks. The multihead LSTM model is comprised of four LSTM models that digest time series data of soil moisture aggregated at different scales as inputs. The outputs of these models, i.e., predicted soil moisture at a certain time in future, are then combined using a weighted averaging method to obtain the final prediction values. Different statistical measures, such as the root mean square error (RMSE), mean square error (MSE), mean absolute error (MAE), mean absolute percentage error (MAPE), and R-squared are employed to assess the performance of the multihead LSTM model predictions against the ground truth. The results show that the proposed multihead LSTM method is effective to forecast the soil moisture up to one-month in future with a R-squared value of 95.04%.