Convolutional Neural Networks accurately predict soil matric potential from soil, weather, and satellite-derived data

Abstract

Being able to predict soil moisture dynamics offers water managers the possibility to better plan irrigation events and prevent soil moisture deficits from reaching levels that reduce crop production. Machine learning (ML) model predictions can potentially assist farmers in managing irrigation water more efficiently. In this study, we aimed to assess the accuracy of a set of ML models in predicting soil matric potential seven days ahead in gravity-surface irrigated cotton paddocks and evaluate the models’ performance for longer term predictions (14 days). The ML models used past soil moisture, weather, and satellite-derived crop-related data as features for the input parameters. Input data were structured in tuples that were organised following a 20-day ‘window’ approach that ‘slid’ one position forward after each training round. A convolutional neural network (CNN) model outperformed a Long Short-Term Memory, Dense Multilayer Perceptron, and Linear Regression model, the latter of which produced the least accurate predictions. The accuracy of the soil matric potential predictions with the CNN model was stable over time (R² ≥ 0.92 and root mean square deviation ≤ 7.5 kPa). However, less accurate predictions were obtained for a short period after emergence and at crop senescence. This study demonstrates the feasibility of producing accurate predictions of soil matric potential in cotton fields at 0.20 m soil depth with a CNN model, which can be integrated into irrigation decision support systems.