Projection climate extreme using Machine learning techniques

Extreme precipitation events are increasing in frequency and intensity in East Africa due to climate change, posing serious threats to agriculture, infrastructure, and human livelihoods. Traditional climate models often struggle to capture the nonlinear, complex, and localized patterns of these extremes. This study explores the use of machine learning specifically Long Short-Term Memory (LSTM) neural networks to improve the projection of extreme precipitation using Coupled Model Intercomparaison project phase 6(CMIP6) climate data under two future scenarios: SSP2-4.5 and SSP5-8.5. To address biases in raw climate model outputs, the Quantile Delta Mapping (QDM) method was applied for bias correction. The corrected data were then used to train and validate LSTM models focused on five major East African cities. Key extreme precipitation indices PRCPTOT, Rx5day, SDII, R95pTOT, and SDII95 were calculated to assess projected changes. Results demonstrate that the combined QDM-LSTM approach significantly improves the accuracy of precipitation projections, capturing both seasonal patterns and long-term trends. Projections indicate an increase in the intensity and variability of extreme rainfall, especially under the high emission SSP5-8.5 scenario. These findings underscore the potential of deep learning models in climate research and highlight their value in supporting climate adaptation strategies and disaster risk management in East Africa.