Scholarlink

<p>Reference evapotranspiration (<i>ET</i><sub><i>o</i></sub>) is a key variable for irrigation scheduling, drought preparedness, and climate-resilient water allocation in semi-arid agriculture. This study developed Bayesian-optimized gradient boosting models to estimate daily <i>ET</i><sub><i>o</i></sub> for two climatically contrasting California stations, Parlier and Meloland, using maximum air temperature, minimum air temperature, relative humidity, and wind speed under RCP 4.5 and RCP 8.5. Historical meteorological inputs were obtained from observed CIMIS records, and benchmark daily <i>ET</i><sub><i>o</i></sub> values used for model development were calculated using the FAO-56 Penman–Monteith equation. Future climate predictors were derived from RCA4/CORDEX-NAM projections, bias-corrected using the delta-change approach, and linked to the corresponding workflow-derived daily ETo targets for scenario-based modeling. Three models were compared: Bayesian-optimized XGBoost (BO-XGB), LightGBM (BO-LGB), and CatBoost (BO-CGB). Cross-validation and independent testing indicated that no single model performed best across all sites and scenarios. On testing, the best RMSE values were 0.129&#xa0;mm/day for BO-XGB at Parlier under RCP 4.5, 0.253&#xa0;mm/day for BO-CGB at Parlier under RCP 8.5, 0.302&#xa0;mm/day for BO-LGB at Meloland under RCP 4.5, and 0.248&#xa0;mm/day for BO-XGB at Meloland under RCP 8.5. Rank aggregation across seven evaluation metrics nevertheless identified BO-XGB as the most consistent overall model. SHAP analysis revealed climate-specific controls on <i>ET</i><sub><i>o</i></sub>, with relative humidity dominating in Parlier, whereas maximum air temperature, followed by wind speed, dominated in Meloland. The study also translated the developed framework into desktop- and web-based GUI tools for rapid scenario-based estimation. Overall, the results provide an interpretable and operational climate-scenario-based framework for <i>ET</i><sub><i>o</i></sub> prediction in California-like semi-arid environments.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Modeling climate change effects on reference evapotranspiration in semi-arid regions using explainable bayesian-optimized gradient boosting models

  • 期刊论文

  • 发表日期：

    2026年7月6日

查看全文
  • Mohamed Kamel Elshaarawy,
  • Romysaa Elasbah,
  • Mohamed Elsayed Gabr,
  • Khaled M. Bali,
  • Mohamed Galal Eltarabily

摘要

Reference evapotranspiration (ETo) is a key variable for irrigation scheduling, drought preparedness, and climate-resilient water allocation in semi-arid agriculture. This study developed Bayesian-optimized gradient boosting models to estimate daily ETo for two climatically contrasting California stations, Parlier and Meloland, using maximum air temperature, minimum air temperature, relative humidity, and wind speed under RCP 4.5 and RCP 8.5. Historical meteorological inputs were obtained from observed CIMIS records, and benchmark daily ETo values used for model development were calculated using the FAO-56 Penman–Monteith equation. Future climate predictors were derived from RCA4/CORDEX-NAM projections, bias-corrected using the delta-change approach, and linked to the corresponding workflow-derived daily ETo targets for scenario-based modeling. Three models were compared: Bayesian-optimized XGBoost (BO-XGB), LightGBM (BO-LGB), and CatBoost (BO-CGB). Cross-validation and independent testing indicated that no single model performed best across all sites and scenarios. On testing, the best RMSE values were 0.129 mm/day for BO-XGB at Parlier under RCP 4.5, 0.253 mm/day for BO-CGB at Parlier under RCP 8.5, 0.302 mm/day for BO-LGB at Meloland under RCP 4.5, and 0.248 mm/day for BO-XGB at Meloland under RCP 8.5. Rank aggregation across seven evaluation metrics nevertheless identified BO-XGB as the most consistent overall model. SHAP analysis revealed climate-specific controls on ETo, with relative humidity dominating in Parlier, whereas maximum air temperature, followed by wind speed, dominated in Meloland. The study also translated the developed framework into desktop- and web-based GUI tools for rapid scenario-based estimation. Overall, the results provide an interpretable and operational climate-scenario-based framework for ETo prediction in California-like semi-arid environments.