A hybrid CNN-GRU model for predicting soil moisture in maize root zone

Abstract

Soil water content in maize root zone is the main basis of irrigation decision-making. Therefore, it is important to predict the soil water content at different depths in maize root zone for rational agricultural irrigation. This study proposed a hybrid convolutional neural network-gated recurrent unit (CNN-GRU) integrated deep learning model that combines a CNN with strong feature expression capacity and a GRU neural network with strong memory capacity. The model was trained and tested with the soil water content and meteorological data from five representative sites in key maize producing areas, Shandong Province, China. We designed the model structure and selected the input variables based on a Pearson correlation analysis and soil water content autocorrelation analysis. The results showed that the hybrid CNN-GRU model performed better than the independent CNN or GRU model with respect to prediction accuracy and convergence rate. The average mean squared error (MSE), mean absolute error and root mean squared error of the hybrid CNN-GRU model on day 3 were 0.91, 0.51 and 0.93, respectively. The prediction accuracy of the model improved with increasing soil depth. Extending the forecast period, the prediction accuracy values of the hybrid CNN-GRU model for the soil water content on days 5, 7 and 10 were comparable, with an average MSE of less than 1.0.