Deep learning for statistical downscaling of daily mean temperature over northwestern algeria using CMIP6 data

Accurate and high-resolution regional climate projections are essential for effective water resource management and climate adaptation, particularly in semi-arid regions such as northwestern Algeria. In this study, we develop a region-specific deep learning (DL) framework based on a convolutional neural network (CNN) to statistically downscale daily mean temperature using CMIP6 large-scale climate predictors. The framework incorporates carefully selected predictors, including topography, distance to the sea, and key synoptic-scale ERA5 variables, and employs a robust training strategy based on the Huber loss function to enhance stability in the presence of occasional extreme temperature values. Results demonstrate that even a relatively compact CNN consistently outperforms a baseline multiple linear regression (MLR) model, with the regional explained variance ([Formula: see text]) increasing from 0.8407 to 0.9116, RMSE decreasing from 4.059 ∘ C to 3.339 ∘ C, and MAE reduced from 3.269 ∘ C to 2.683 ∘ C. These improvements highlight the CNNs ability to capture physically coherent relationships between local physiographic factors and large-scale atmospheric predictors, producing a more accurate spatial representation of temperature variability. Additional diagnostics further confirm the robustness of the proposed framework in reproducing both mean climate characteristics and spatial error structures. Overall, the study demonstrates the potential of DL as a powerful and transferable tool for statistical climate downscaling over Algeria and similar climate-sensitive regions.