Statistical Downscaling of Global Climate Models for Temperature and Precipitation over West Africa Using Machine Learning Algorithms

Abstract

Global Climate Models (GCMs) provide essential projections for understanding future climate conditions but operate at coarse spatial resolutions, limiting their applicability for regional and local-scale climate impact assessments. To address this gap, this study applies advanced machine learning (ML) based statistical downscaling techniques to improve the spatial resolution and accuracy of temperature and precipitation projections over West Africa. Using historical climate observations (1985–2014) and future CMIP6 GCM outputs (2015–2100), we implement and evaluate both regression-based and deep learning (DL) models specifically, Convolutional Neural Networks (CNNs) and Convolutional Long Short-Term Memory networks (ConvLSTM). The study assesses downscaling performance across multiple GCMs (CMCC-ESM2, CMCC-CM2-SR5, FGOALS-g3, NorESM2-LM), comparing model outputs against high-resolution reference datasets (ERA5 for temperature and CHIRPS for precipitation). Results reveal that CNN models are particularly effective in downscaling precipitation, offering high spatial fidelity, reduced biases, and superior accuracy (RMSE = 0.137 mm/day, R² = 0.995), especially in humid and orographic zones. Conversely, ConvLSTM models outperform CNNs in temperature downscaling, capturing temporal dependencies and seasonality with high precision (RMSE = 0.212 °C, R² = 0.990, r = 0.996). While both DL models significantly outperform raw GCM outputs, CNN is better suited for applications requiring spatial detail, such as rainfall intensity mapping, while LSTM is more appropriate for time-sensitive applications, including temperature forecasting and hydrological modeling. This research confirms the potential of deep learning–based statistical downscaling to enhance the usability of GCM outputs for regional climate analysis and adaptation planning in West Africa. The results contribute to ongoing efforts to build robust, high-resolution climate datasets that can support informed policy decisions and sectoral planning under climate change.