Machine-learning prediction of dust aerosol optical depth at 550 nm over West Africa

Abstract Dust aerosol optical depth (AOD) over West Africa is controlled by interacting meteorological, land-surface and seasonal factors. We have developed a machine-learning framework to predict daily dust AOD at 550 nm wavelength over West Africa from 2003 to 2024 using meteorological and land-surface reanalysis variables. Linear Regression, Random Forest and LightGBM were compared using a time-based split for training, validation, and independent test periods. LightGBM gave the strongest validation performance and was therefore selected for independent testing. The model reproduced the north-south dust-AOD gradient and the seasonal structure, although errors remained larger during March-May and June-August. Probability density and cumulative distribution analyses showed that LightGBM captured much of the low-to-moderate dust AOD range and predicted fewer high-dust events. At the 90 th percentile, the model detected 41% of high-dust events and produced a frequency bias of 0.76. SHapley Additive exPlanations (SHAP) analysis identified temperature, wind speed, relative humidity and soil moisture as the leading predictors. The spatial self-organising map further showed that the low-to-moderate dust patterns were better represented than the high dust patterns. The results define both the capability and the remaining limitations of the dust AOD predicting framework over West Africa.