Soil Moisture Prediction for Intelligent Irrigation: An XGBoost-based Model with Multi-Dimensional Feature Engineering

Abstract

Accurate soil moisture prediction is essential for intelligent irrigation and optimal agricultural water resource allocation. To address the limitation of traditional linear models in capturing complex nonlinear relationships between meteorological factors and soil moisture, this study developed an Extreme Gradient Boosting (XGBoost)-based soil moisture prediction model using hourly meteorological and soil moisture data from a 1-hectare multi-crop farm, achieving high-precision prediction through in-depth data preprocessing, multi-dimensional feature engineering, and systematic model training and validation. Results indicate that with 44 optimal features selected, the model achieves a coefficient of determination (R²) of 0.673 on the test set, with low mean squared error (MSE) and mean absolute error (MAE), outperforming traditional linear models significantly; feature importance analysis identifies daily average temperature, previous day's soil moisture, and total daily precipitation as key driving factors (consistent with soil water evaporation and replenishment mechanisms); and the model predicts 5 cm depth soil moisture of 0.2435 (meeting the minimum crop survival threshold) for the target date. This model provides reliable data support for dynamic decision-making in intelligent irrigation systems, with great practical value for improving agricultural water use efficiency and advancing precision agriculture.