This study examines the bias corrections (BCs) to ERA5-reanalyzed daily maximum (Tmax) and minimum (Tmin) temperatures for the metropolitan environments of western India, located on the coastal plain (Mumbai), elevated plateau landmass (Pune), and inland plain (Ahmedabad), using quantile mapping (QM) and machine learning (ML) approaches. Four ML models are used in this study that are trained with 18 predictors against the observed set of Tmax and Tmin. ERA5-diagnoses based on the period 1980–2019 in these environments exhibit cold [warm] bias in Tmax [Tmin], with the largest bias seen in the coastal environment. Application of QM generally shows greater skill in BCs of ERA5 temperature distributions, and substantially reduces RMSEs and improves the reproduction of observed variability in the coastal and plateau environments on annual and seasonal scales. All ML models used in this study showed near-similar performances in BCs and outperformed the QM BCs. The ML approaches show the largest BC skill of 80% [60%] in Tmax [Tmin] in the coastal environment and also exhibit 20–30% higher BC skill achieved by QM across all environments. The ML-driven BCs also display 3–4 [2] fold RMSE reductions in Tmax [Tmin] in all environments relative to ERA5 fields, and enhanced retrieval of observed variability in Tmax in the coastal environment relative to QM. They also show greater skill in exhibiting the inter-annual variations and interdependence in Tmax and Tmin than QM. The skill diagnostics further show ML and QM-corrected Tmax events in the range 300–310 K [300–315 K] have undergone significant BCs in the coastal [elevated plateau] environment. While QM and ML substantially improve the representation of the upper tail of the raw-ERA5 diagnosed Tmax distribution, ML methods offer additional improvements ranging from 15% to 36% in skill across different environments. Furthermore, ML-based bias correction improves the accuracy of heat wave diagnoses by 27–40% compared to raw-ERA5 across different environments. This investigation highlights the relative advantages of less computationally expensive data-driven approaches and renders promising insights into the ML application with physics-aware predictors set towards better benefits of identifying the regional hotspots of extreme heat episodes from reanalyzed products. While this study is limited in spatial scope, evaluating the generalizability of the model across regions remains an important direction for future research.

Bias corrections in reanalyzed daily maximum and minimum temperatures for the metropolitan environments in Western India

期刊论文
发表日期：

2026年3月25日

查看全文

Shwetha Rajashree Iyer,
Shubham Prajapati,
Mariya Joy Kurien Reji,
Sumita Kedia,
Bagati Sudarshan Patro,
Ramesh Kumaraswamy Vellore,
Bhupendra Bahadur Singh

摘要

This study examines the bias corrections (BCs) to ERA5-reanalyzed daily maximum (T_max) and minimum (T_min) temperatures for the metropolitan environments of western India, located on the coastal plain (Mumbai), elevated plateau landmass (Pune), and inland plain (Ahmedabad), using quantile mapping (QM) and machine learning (ML) approaches. Four ML models are used in this study that are trained with 18 predictors against the observed set of T_max and T_min. ERA5-diagnoses based on the period 1980–2019 in these environments exhibit cold [warm] bias in T_max [T_min], with the largest bias seen in the coastal environment. Application of QM generally shows greater skill in BCs of ERA5 temperature distributions, and substantially reduces RMSEs and improves the reproduction of observed variability in the coastal and plateau environments on annual and seasonal scales. All ML models used in this study showed near-similar performances in BCs and outperformed the QM BCs. The ML approaches show the largest BC skill of 80% [60%] in T_max [T_min] in the coastal environment and also exhibit 20–30% higher BC skill achieved by QM across all environments. The ML-driven BCs also display 3–4 [2] fold RMSE reductions in T_max [T_min] in all environments relative to ERA5 fields, and enhanced retrieval of observed variability in T_max in the coastal environment relative to QM. They also show greater skill in exhibiting the inter-annual variations and interdependence in T_max and T_min than QM. The skill diagnostics further show ML and QM-corrected T_max events in the range 300–310 K [300–315 K] have undergone significant BCs in the coastal [elevated plateau] environment. While QM and ML substantially improve the representation of the upper tail of the raw-ERA5 diagnosed T_max distribution, ML methods offer additional improvements ranging from 15% to 36% in skill across different environments. Furthermore, ML-based bias correction improves the accuracy of heat wave diagnoses by 27–40% compared to raw-ERA5 across different environments. This investigation highlights the relative advantages of less computationally expensive data-driven approaches and renders promising insights into the ML application with physics-aware predictors set towards better benefits of identifying the regional hotspots of extreme heat episodes from reanalyzed products. While this study is limited in spatial scope, evaluating the generalizability of the model across regions remains an important direction for future research.