Soil Moisture Prediction Based on Long Short-Term Memory Networks and Meteorological Data

As an important factor directly affecting the growth of crops, a reasonable and accurate prediction of soil moisture can effectively improve the quality of crop growth. The deep learning model LSTM is often used to predict soil moisture, nevertheless, the predictions generated only by LSTM exhibit a lack of accuracy. To resolve this issue, in this study, we propose a method that combines meteorological features and correlation of soil at different depths to predict soil moisture. By verifying the autocorrelation and cross-correlation of soil at different depths, it is determined that the soil moisture characteristics can be regarded as a smooth time series, and the meteorological data can be used to predict the soil moisture. Afterwards, the LSTM model was used to predict soil moisture from meteorological data, and it was found that there was a significant disparity between the actual value and the predicted value, and the accuracy of prediction was improved by reconstructing the model. To validate the proposed model, we selected moisture data from six monitoring points in unirrigated (rainfed) wheat fields in the designated areas of the Yellow and Huaihai Seas to validate the proposed model, and obtained accurate moisture prediction values, proving the validity of the model.