Improving soil moisture prediction using Gaussian process regression

Abstract

Soil moisture plays a vital role in agriculture and hydrology, influencing key processes like plant growth and evaporation. Recent advancements in soil moisture monitoring have improved our ability to measure it at different scales, but challenges persist at intermediate scales that are crucial for precision agriculture. To address this research gap, innovative methods like machine learning (ML) are being explored to improve prediction accuracy, overcoming the limitations of traditional models. By leveraging an extensive dataset that spans multiple sites and seasons, we aim to improve predictions for both surface and root zone soil moisture. In this study, machine learning models including multilinear regression (MLR), support vector machine (SVM), and Gaussian process regression (GPR), were developed and compared for soil moisture predictions at different depths at 29 weather stations in the Red River Valley using features such as time, locations, meteorological data, soil physical properties, and remote sensing data. Our research showed that GPR with automatic relevant determination kernels had the best performance with $R^2$ values greater than 0.95 at almost all depths when including all features. GPR ($R^2$=0.95–0.99, RMSE=0.0045–0.0224, MAE=0.0012–0.0139) outperformed MLR ($R^2$=0.69–0.93, RMSE=0.0328–0.0555, MAE=0.0197–0.0427) and SVM ($R^2$=0.49–0.85, RMSE=0.0648–0.0747, MAE=0.0442–0.0566) for soil moisture prediction. All models performed better when predicting moisture in subsoils (20–100 cm) than in topsoil (0–10 cm). Our research highlights the effectiveness of GPR as a powerful ML tool that enhances soil moisture management precision, ultimately contributing to more effective and smart agricultural practices.