Ensemble Machine Learning Approaches for Statistical Downscaling and Future Precipitation Projection in Maiduguri, Nigeria

Abstract

Climate change impact studies are carried out using the general circulation models (GCMs) to mitigate the greenhouse gas effect on the environment. This study aimed to determine the long-term future changes of precipitation in Maiduguri Nigeria. To achieve this, the GCM variables were downloaded in coarse resolution and converted into local scale using machine learning (ML) and conventional downscaling models including booted regression trees (BOT), support vector machine (SVM) and multiple linear regression (MLR) with prior selection of dominant inputs by Pearson correlation analysis. To improve the downscaling performance of the standalone models, ensemble approach was applied. The ensemble approach has the ability to combine the strength and weaknesses of the single models thereby improving performance. Thereafter, BOT and SVM models were applied to forecast the future precipitation changes. The results showed that with appropriate selection of the GCM variables, BOT, SVM and MLR models could be employed for the statistical downscaling in Maiduguri station. Moreover, the applied ensemble model has improved performance up to 8%, 29% and 30% over BOT, SVM and MLR respectively. The forecast results indicated a decrease in precipitation amount most especially in rainy season months from June to September towards the year 2092.