Comparison of conventional and machine learning methods for bias correcting CMIP6 rainfall and temperature in Nigeria

This research assesses the efficacy of thirteen bias correction methods, including traditional and machine learning-based approaches, in downscaling four chosen GCMs of Coupled Model Intercomparison Project 6 (CMIP6) in Nigeria. The 0.5° resolution gridded rainfall, maximum temperature (Tmx), and minimum temperature (Tmn) of the Climate Research Unit (CRU) for the period 1975 − 2014 was used as the reference. The Compromise Programming Index (CPI) was used to assess the performance of bias correction methods based on three statistical metrics. The optimal bias-correction technique was employed to rectify bias to project the spatiotemporal variations in rainfall, Tmx, and Tmn over Nigeria for two distinct future timeframes: the near future (2021–2059) and the distant future (2060–2099). The study's findings indicate that the Random Forest (RF) machine learning technique better corrects the bias of all three climate variables for the chosen GCMs. The CPI of RF for rainfall, Tmx, and Tmn were 0.62, 0.0, and 0.0, followed by the Power Transformation approach with CPI of 0.74, 0.36, and 0.29, respectively. The geographic distribution of rainfall and temperatures significantly improved compared to the original GCMs using RF. The mean bias-corrected projections from the multimodel ensemble of the GCMs indicated a rainfall increase in the near future, particularly in the north by 2.7–12.7%, while a reduction in the south in the far future by -3.3% to -10% for different SSPs. The temperature projections indicated a rise in the Tmx and Tm from 0.71 °C and 0.63 °C for SSP126 to 2.71 °C and 3.13 °C for SSP585. This work highlights the significance of comparing bias correction approaches to determine the most suitable approach for adjusting biases in GCM estimations for climate change research.