Time-Series Analysis and Learning Approaches for Soil Moisture Prediction in Precision Irrigation
Please login to view abstract download link
Soil water content is a key element for managing precision irrigation in crops, especially in agricultural contexts where water stress and soil salinity pose significant challenges, thus motivating the interest in deepening the forecast of this variable. This study examines high-frequency monitoring data collected from an experimental agricultural site in Gallipoli, Italy. The dataset includes time series of soil water content measurements at different depths, atmospherical data (mainly precipiation and temperature) irrigation data, with different temporal resolution. We compare a range of data-driven forecasting models, including traditional statistical approaches such as Holt-Winters and VARIMA in R language, machine learning techniques like Random Forest and Gradient Boosting regression, and deep learning architectures exemplified by Chronos. To reflect the diverse nature of the forecasting tasks considered we integrate external variables, particularly irrigation data, along with precipitation and weather conditions, enhancing the models’ ability to capture the real dynamics of water content in the soil. By comparing univariate and multivariate approaches, we demonstrate that incorporating external data enables multivariate models to achieve more accurate and reliable short-term predictions. The integration of data from different platforms makes our methodology a practical tool for forecasting soil water content and therefore optimizing water use and promoting more sustainable agricultural management.