A hybrid machine learning-based past performance and envelope approach for rainfall projection in Sarawak, Malaysia

Abstract

This study assesses historical and future rainfall patterns in Sarawak's diverse ecology by using a novel hybrid machine learning-based past-performance and envelope approaches to select the most suitable global climate models (GCMs) for climate projections. Additionally, a non-local Model Output Statistics (MOS) approach is introduced for climate downscaling, enhancing the precision of localized projections. A frequency-based approach identified the optimal GCMs (HadGEM2-AO, HadGEM2-ES, CCSM4, and CESM1-CAM5), effectively refining the selection of models required for accurate and reliable rainfall projections. The Support Vector Machine (SVM)-based downscaling models developed were able to replicate historical rainfall with a mean error below 50 mm/month. Seasonal changes were most pronounced in January, with increases ranging from 1.8 % to 32.9 %, except in December under RCP8.5 during 2040–2069, which showed the highest increase of 9.0 %. The most notable rainfall decrease occurred in July, ranging from −16.4 % to −38.9 %. Increased rainfall during the peak months of the Northeast Monsoon (NEM) indicates a heightened concentration of rainfall, which could contribute to more frequent hydro-climatological extremes. Conversely, decreased rainfall is projected for other NEM months (February, March, November) as well as throughout April to October, suggesting an increased likelihood of prolonged dry periods during the Southwest Monsoon (SWM) in the future. These findings underscore the importance of the study's hybrid machine learning-driven GCM selection and non-local MOS downscaling method in improving rainfall projections for Sarawak, providing high-resolution data to support government climate resilience efforts and policy development.