Climate-Driven Multi-Depth Soil Moisture and Irrigation Demand Projections Using Bias- Corrected CMIP6 and Machine Learning in the Amibara Agrucultural Land Near Kulfo Watershed, Ethiopia
Year: 2026
Type: preprint
Abstract: Climate change can modify precipitation regimes, atmospheric evaporative demand, and soil-water availability, creating substantial challenges for irrigation-dependent agriculture. Soil moisture responds differently across the soil profile, assessments based on a single soil layer may obscure important depth-dependent responses. This study developed an integrated framework combining bias-adjusted CMIP6 climate projections, machine learning, explainable artificial intelligence, recursive forecasting, and root-zone water-balance modelling to investigate climate-driven changes in multi-depth soil moisture and irrigation requirements in the Kulfo watershed, southern Ethiopia. Daily precipitation, maximum temperature, minimum temperature, and soil-moisture observations at 10, 20, 30, and 40 cm were analyzed for 2006–2014. Eighteen CMIP6 general circulation models (GCMs) were evaluated and bias-adjusted using a two-stage Quantile Mapping–Random Forest (QM–RF) framework. The analysis separated the historical period chronologically into 2006–2010 for model development, 2011–2012 for validation and model selection, and 2013–2014 for independent testing. Five machine-learning approaches of CatBoost, Histogram-Based Gradient Boosting, LightGBM, XGBoost, and a stacking ensemble were evaluated using precipitation, temperature, antecedent precipitation, lagged temperature, and lagged soil-moisture predictors. Because lagged soil moisture is unavailable as an observed quantity during future simulation, recursive hindcasting was used to evaluate the effect of feeding model predictions back into subsequent time steps. SHapley Additive exPlanations (SHAP) were used to quantify predictor contributions to soil-moisture predictions. Following bias adjustment, CMCC-ESM2 showed the highest precipitation performance index (PI = 0.83), whereas NorESM2-MM showed the highest indices for maximum and minimum temperature (PI = 0.87 and 0.85, respectively). QM–RF reduced precipitation RMSE by 16–22% relative to QM alone (mean 18.1%). CatBoost achieved the strongest independent-test performance at 40 cm (R² = 0.918; RMSE = 0.0134 cm³/cm³), while HistGB was best at 20 cm (R² = 0.956). Under recursive hindcasting over the independent 2013–2014 period, R² values were 0.861, 0.901, 0.889, and 0.854 at 10, 20, 30, and 40 cm, respectively. Under SSP5-8.5, annual mean soil moisture was projected to decline by approximately 29%, 27%, 24%, and 21% at 10, 20, 30, and 40 cm, respectively, relative to the historical baseline. The 10th–90th percentile inter-model spread for these projections ranged from ± 6% to ± 11%. Crop water requirement increased from 620 mm during the historical baseline to 714 mm during 2071–2100 under SSP5-8.5, while net irrigation requirement increased from 210 to 301 mm. The combined reduction in soil-water availability and increase in crop-water demand indicates potentially greater irrigation pressure under stronger climate forcing. However, the projected changes should be interpreted as conditional scenario simulations rather than deterministic forecasts because they depend on climate-model structural uncertainty, bias-adjustment assumptions, recursive machine-learning behavior, the relatively short observational period, and crop-water-balance parameterization. Explicit representation of inter-model spread is therefore required when interpreting the magnitude and robustness of the future projections.
Source: Research Square
Institution: Arba Minch University
Language: English
Subfield: Global and Planetary Change
Field: Environmental Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7225589892