Abstract
Accurately predicting particulate matter is crucial for preventing health risks and protecting public health. This study improves the accuracy of particulate matter d ≤ 10μm (PM10) forecasts over Morocco for the next five days using a U-Net-based deep learning model, marking the first work of its kind in the Middle East and North Africa (MENA) region. The U-Net model was used to post-process and improve PM10 forecasts from the Copernicus Atmosphere Monitoring Service (CAMS), with reanalysis data from CAMS serving as a reference to guide the model's learning. The U-Net architecture was modified to predict outputs at a resolution different from the inputs, eliminating the need for interpolation and preserving critical spatial details. The results demonstrated significant improvements over two baselines-CAMS forecasts and the Analog Ensemble model (AnEn)-by enhancing metrics such as Mean Absolute Error (MAE), Root Mean Square Error (RMSE), Coefficient of Determination ( R2 ), Index of Agreement (IOA), and biases, particularly in regions prone to dust storms, during the period prior to the CAMS forecast upgrade in mid-2023. In the second half of 2023, U-Net continued to improve predictions; however, the effect of the upgrade cycle became evident in its errors. This highlights the importance of retraining U-Net with updated data as it becomes available to maintain its reliability in operational forecasting systems. U-Net also proved effective in capturing particulate pollution, providing reliable predictions for values up to 500 μg/m3 . These findings underscore U-Net's potential for operational forecasting, supporting accurate early warnings to mitigate the health and environmental impacts of PM10 pollution.