Recurrent Neural Networks for Soil Moisture Prediction Leveraging Soil Matric Potential Data

Soil moisture is a parameter of paramount importance for a variety of applications, such as predicting floods and droughts, monitoring agricultural crop performance, and managing water supply. It can be measured in several ways, such as Volumetric Water Content (VWC) and soil matric potential. An experiment was carried out by placing soil matric potential sensors at depths of 20 centimeters and 40 centimeters within the root layer of an adult apple tree orchard to gather data every 10 minutes and to train some Recurrent Neural Network (RNN) models to predict future matric potential values at the depths of interest. Base RNN, Gated Recurrent Unit (GRU), and Long Short-Term Memory (LSTM) networks were employed to accomplish the aforementioned goal. Additionally, feature selection analysis was used to determine the best parameters to feed the models. The trained models give an accurate short-term prediction of 10 minutes, with an R² of 0.9947, and a long-term prediction of 3 hours, with an R² of 0.7922 at 20 centimeters.