Time–Series Analysis and Learning Approaches for Soil Moisture Prediction in Precision Irrigation: Dataset and Code

This repository contains the data and code associated with the study on forecasting soil water content for precision irrigation purposes. The work is based on high-frequency monitoring data collected at an experimental agricultural field located in Gallipoli, southern Italy. The dataset includes: Soil water content measurements every 15 minutes from six sensor ports, with odd-numbered ports installed at 30 cm depth and even-numbered ports at 60 cm depth. Rainfall Irrigation events with varying temporal resolution. The code implements and compares several forecasting approaches: A traditional statistical model: Holt-Winters (in R), Machine learning regressors: Random Forest, Gradient Boosting (in Python), A deep learning model for time series prediction: Chronos (in Python). External variables such as rainfall and irrigation are integrated to improve model performance, with a comparison between univariate and multivariate strategies. The forecasts were evaluated over prediction horizons ranging from 48 hours to 5 days, including intermediate 2-day forecasts, under conditions characterized by high temporal variability in soil moisture dynamics. The goal is to support short-term forecasting of soil water content in real-world agricultural settings, contributing to optimized irrigation and sustainable water resource management.