A Comparative Analysis of Machine Learning Regression Algorithms for Soil Moisture Estimation
Soil moisture is a fundamental hydrological state variable that governs the exchange of water and energy between the land surface and the atmosphere, and it has direct implications for agricultural productivity, drought monitoring and ecosystem management. This study benchmarks thirteen machine learning regressors for soil moisture forecasting on a long-term hourly record from the University of Toronto Mississauga campus, and asks two questions that comparative studies of this kind usually leave implicit: at what forecast horizon does machine learning actually add value over a naive baseline, and how much of the accuracy reported in the literature is an artefact of the evaluation protocol? All models are trained on 2009–2016 and evaluated on a strictly subsequent hold-out period covering September 2016 to April 2018, with hyperparameters selected by randomised search under blocked time-series cross-validation restricted to the training period. Forecast horizons from 1 to 168 hours are examined. At the one-hour horizon persistence is already almost perfect (R²=0.9918) and machine learning adds nothing measurable; the advantage becomes positive from three hours onwards and grows monotonically, reaching +0.044 at 24 hours and +0.260 at one week, where persistence collapses to R²=0.014. At the operationally relevant 24-hour horizon the best model, XGBoost, attains R²=0.8033, MAE=0.0152 m³/m³ and RMSE=0.0252 m³/m³ against a persistence baseline of R²=0.7589, and a Holm-corrected Wilcoxon signed-rank test confirms the improvement (P=3.6×10⁻⁴²). A Friedman test establishes systematic differences among the algorithms (χ²=31.44, P=1.69×10⁻³), yet XGBoost cannot be statistically separated from ordinary linear regression, which reaches R²=0.7971 at roughly one-seventieth of the training cost. A feature ablation shows why: soil moisture history supplies almost all of the predictive signal at short horizons, while the atmospheric variables contribute measurably only from twelve hours onwards. Finally, replacing the chronological split by a random one inflates R² by a median of 8.2% when history features are present, but by 65.4% when they are not, showing that the severity of temporal leakage depends on the feature set as well as on the partitioning scheme.
Authors
- Erdener Özçetin (ORCID: https://orcid.org/0000-0002-6079-3159)
Institutions
- Hitit Üniversitesi (TR)
Publication Details
- Journal
- Black Sea Journal of Agriculture
- Published
- 2026-11-15
- DOI
- https://doi.org/10.47115/bsagriculture.1946130
- Primary Topic
- Hydrological Forecasting Using AI
- Type
- article
- Field-Weighted Citation Impact
- 0.00