Developing Machine Learning and Deep Learning Soil Moisture Models for Precision Agricultural Applications

Abstract

Monitoring soil moisture is increasingly becoming a research focus in the fields of agriculture, hydrology, meteorology, and ecology. While soil moisture measurements at points (<1 m²) and their estimation at larger scales (100-25,000 km²) have improved considerably, soil moisture modeling at the intermediate scales (10 to 100 m²) needs more attention. In this study, machine learning and deep learning models including multi-linear regression (MLR), support vector machine (SVM), Gaussian process regression (GPR), and convolutional neural networks (CNN) were built and compared for soil moisture predictions at different depths at the weather stations in the Red River Valley using locations, meteorological data, and soil physical properties. The results showed that the GPR (R²=0.80-0.90) outperformed other models including MLR (R²=68-0.82), SVM (R²=0.44-0.60), and CNN (R²=0.66-0.84) for soil moisture prediction. The prediction performance in the topsoil was better than in the subsoils. The GPR outperformed SVM when both models used the same kernel functions and kernel parameters.