PREDICTION OF SOIL MOISTURE BY OPTIMIZABLE ENSEMBLE MACHINE LEARNING APPROACH USING SUB HOURLY WEATHER PARAMETERS

Abstract

Soil moisture influences seed germination, nutrient transport, microbial activity and crop productivity and its accurate prediction is becoming more important as rainfall patterns grow more irregular and temperatures rise. This study evaluated regression-based machine learning models for predicting soil moisture from meteorological data, including air temperature, relative humidity, wind speed, sunshine duration and soil temperature, recorded at an Automatic Weather Station at 15-minute intervals between October 2021 and November 2023. Twenty-eight sub-models spanning eight machine learning categories were compared using MAE, MSE, RMSE, RSE and R². An optimizable ensemble model tuned with grid search gave the best validation performance, with an MAE of 1.9756, MSE of 8.4283, RMSE of 2.9031 and R² of 0.83 and test performance of MAE 1.8240, MSE 7.5353, RMSE 2.7451 and R² of 0.85. Permutation importance and SHAP (Shapley Additive Explanations) analyses both ranked relative humidity, soil temperature and time of observation as the most influential predictors and SHAP dependence plots showed relationships between these variables and the predicted soil moisture that were consistent with known evaporative and diurnal processes. The model achieved this performance using only above ground meteorological inputs, without soil moisture as a predictor. These findings indicate that an optimized ensemble model can predict soil moisture from routinely measured weather variables with reasonable accuracy and interpretability, providing a potential basis for future irrigation-management and drought-monitoring applications.