High-resolution machine-learning downscaled climate projections and extreme hazard assessment over monsoon-dominated Odisha, India

This study develops a grid-wise eXtreme Gradient Boosting (XGBoost) framework for bias correction and statistical downscaling of daily precipitation, maximum temperature (Tmax), and minimum temperature (Tmin) over Odisha using 30 years (1985–2014) of India Meteorological Department (IMD) gridded observations at 0.25° spatial resolution. The framework was applied to simulations from CORDEX-SA–CMIP5 and NEX-GDDP–CMIP6 datasets. Model evaluation employed an Equal-Weight Composite Score, Monte Carlo stability analysis, and split-period transferability testing to assess robustness and temporal consistency. CORDEX-SA demonstrated stable temporal transferability, whereas NEX-GDDP–CMIP6 precipitation exhibited poor stationarity associated with embedded BCSD double-correction artefacts; therefore, CORDEX-SA was selected for future climate projections. XGBoost substantially reduced systematic biases, decreasing CORDEX-SA precipitation RMSE by 58.6% and Tmax RMSE by 95.7%. Comparison with Quantile Mapping (QM) showed that XGBoost generally outperformed QM for mean-state correction, while QM better preserved upper-tail precipitation extremes. Future projections under RCP4.5 and RCP8.5 indicate progressive warming across Odisha, with domain-mean Tmax increasing by 1.61°C and 2.68°C, respectively, by 2070–2099, alongside precipitation declines of up to 9.74%. ETCCDI analysis further indicates intensified heat extremes and 6–8 additional consecutive dry days over interior Odisha under RCP8.5.