Scholarlink

<p>This study quantitatively evaluates the performance of CMIP6-based neural emulators and machine learning regressors in forecasting district-level maximum surface air temperature (SAT) across Telangana during the pre-monsoon season (March–May, 1985–2014). Daily SAT data from 27 CMIP6 models and IMD Ensemble observations were harmonized through interpolation, bias correction, and outlier filtering. Three emulator configurations, Multi-Model Mean (MMM_Group-20), Shallow ANN (sANN_Group-20), and Deep ANN (dANN_Group-20), were benchmarked against CMIP6 Ensembles using statistical diagnostics, Taylor metrics, and spatial bias mapping. Results revealed persistent warm biases (+ 1.5 to + 3.5&#xa0;°C), compressed confidence intervals (0.02–0.05&#xa0;°C vs. 0.04–0.12&#xa0;°C in Ensembles), and smoothing of extremes. Taylor analysis indicated RMSE values of 5.09–5.17&#xa0;°C and correlation coefficients of 0.24–0.36, reflecting ensemble-like performance but limited improvements over physics-based models. District-level bias maps highlighted northern hotspots (Mancherial, Adilabad, Jagtial), while reliability scoring identified MMM_Group-20 as the most consistent, achieving scores above 0.85 and peaking at 1.240 under Gradient Boosting. Ensemble regressors such as XGBoost and Random Forest delivered the highest mean reliability (~ 0.465), whereas Linear Regression underperformed in anomaly-dense districts. Overall, the findings demonstrate that neural emulators provide scalable and stable forecasting frameworks but exaggerate heat profiles and compress uncertainty ranges. Post-hoc bias correction and interpretability tools, including SHAP-based feature attribution, are essential to enhance realism. By linking emulator biases to structural constraints and large-scale drivers such as ENSO, this framework advances AI-driven climate diagnostics and supports agro-climatic decision-making in monsoon-affected regions.</p> Graphical Abstract <p></p> <p>This study evaluates CMIP6-based neural emulators and machine learning regressors for predicting district-level maximum surface air temperature (SAT) across Telangana during the pre-monsoon season (March–May, 1985–2014). Daily SAT from 27 CMIP6 models, combined with IMD Ensemble observations, was harmonized through interpolation, bias correction, and outlier filtering. Three emulator strategies were tested: Multi-Model Mean (MMM_Group_20), shallow ANN (sANN_Group_20), and deep ANN (dANN_Group_20), trained via mean squared error minimization and benchmarked against CMIP6 Ensembles. Diagnostics revealed persistent warm biases (+ 1.5 to + 3.5&#xa0;°C), narrower confidence intervals (0.02–0.05&#xa0;°C vs. 0.04–0.12&#xa0;°C in Ensembles), and smoothing of extremes. Taylor diagrams showed ensemble-like performance but limited gains over physics-based models, with RMSE ~ 5.1&#xa0;°C and correlations 0.24–0.36. District-level bias maps highlighted hotspots in northern Telangana (Mancherial, Adilabad, Jagtial). Reliability scoring of five regressors confirmed emulator robustness: ensemble methods (XGBoost, Random Forest) achieved mean reliability ~ 0.465, Gradient Boosting peaked at 1.240, while Linear Regression underperformed in anomaly-prone districts.Findings suggest neural emulators provide scalable, stable forecasting frameworks but exaggerate heat profiles and compress uncertainty, requiring post-hoc bias correction and interpretability tools. Linking emulator biases to large-scale drivers like ENSO strengthens AI-driven climate diagnostics and supports agro-climatic decision-making in monsoon-sensitive regions.</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

Neural Emulators vs. Machine Learning Regressors: District‑Scale Pre‑Monsoon Temperature Prediction in Telangana

  • 期刊论文

  • 发表日期：

    2026年5月14日

查看全文
  • Guhan Velusamy,
  • Dharma Raju Akasapu,
  • Nagaratna Kopparthi,
  • G. Ch. Satyanarayana,
  • Sheshakumar Goroshi,
  • Gamal El Afandi

摘要

This study quantitatively evaluates the performance of CMIP6-based neural emulators and machine learning regressors in forecasting district-level maximum surface air temperature (SAT) across Telangana during the pre-monsoon season (March–May, 1985–2014). Daily SAT data from 27 CMIP6 models and IMD Ensemble observations were harmonized through interpolation, bias correction, and outlier filtering. Three emulator configurations, Multi-Model Mean (MMM_Group-20), Shallow ANN (sANN_Group-20), and Deep ANN (dANN_Group-20), were benchmarked against CMIP6 Ensembles using statistical diagnostics, Taylor metrics, and spatial bias mapping. Results revealed persistent warm biases (+ 1.5 to + 3.5 °C), compressed confidence intervals (0.02–0.05 °C vs. 0.04–0.12 °C in Ensembles), and smoothing of extremes. Taylor analysis indicated RMSE values of 5.09–5.17 °C and correlation coefficients of 0.24–0.36, reflecting ensemble-like performance but limited improvements over physics-based models. District-level bias maps highlighted northern hotspots (Mancherial, Adilabad, Jagtial), while reliability scoring identified MMM_Group-20 as the most consistent, achieving scores above 0.85 and peaking at 1.240 under Gradient Boosting. Ensemble regressors such as XGBoost and Random Forest delivered the highest mean reliability (~ 0.465), whereas Linear Regression underperformed in anomaly-dense districts. Overall, the findings demonstrate that neural emulators provide scalable and stable forecasting frameworks but exaggerate heat profiles and compress uncertainty ranges. Post-hoc bias correction and interpretability tools, including SHAP-based feature attribution, are essential to enhance realism. By linking emulator biases to structural constraints and large-scale drivers such as ENSO, this framework advances AI-driven climate diagnostics and supports agro-climatic decision-making in monsoon-affected regions.

Graphical Abstract

This study evaluates CMIP6-based neural emulators and machine learning regressors for predicting district-level maximum surface air temperature (SAT) across Telangana during the pre-monsoon season (March–May, 1985–2014). Daily SAT from 27 CMIP6 models, combined with IMD Ensemble observations, was harmonized through interpolation, bias correction, and outlier filtering. Three emulator strategies were tested: Multi-Model Mean (MMM_Group_20), shallow ANN (sANN_Group_20), and deep ANN (dANN_Group_20), trained via mean squared error minimization and benchmarked against CMIP6 Ensembles. Diagnostics revealed persistent warm biases (+ 1.5 to + 3.5 °C), narrower confidence intervals (0.02–0.05 °C vs. 0.04–0.12 °C in Ensembles), and smoothing of extremes. Taylor diagrams showed ensemble-like performance but limited gains over physics-based models, with RMSE ~ 5.1 °C and correlations 0.24–0.36. District-level bias maps highlighted hotspots in northern Telangana (Mancherial, Adilabad, Jagtial). Reliability scoring of five regressors confirmed emulator robustness: ensemble methods (XGBoost, Random Forest) achieved mean reliability ~ 0.465, Gradient Boosting peaked at 1.240, while Linear Regression underperformed in anomaly-prone districts.Findings suggest neural emulators provide scalable, stable forecasting frameworks but exaggerate heat profiles and compress uncertainty, requiring post-hoc bias correction and interpretability tools. Linking emulator biases to large-scale drivers like ENSO strengthens AI-driven climate diagnostics and supports agro-climatic decision-making in monsoon-sensitive regions.