High-Resolution Emission Inventory using CNN-LSTM in Congo Basin in Africa. NO2 and PM2.5 Modeling for Kinshasa and Brazzaville

This study addresses critical air quality data gaps in the Congo Basin by developing the first high-resolution anthropogenic emission inventory for the Kinshasa-Brazzaville metropolitan area, a major African megacity with severe pollution but sparse monitoring. We introduce a novel machine-learning framework that fuses multi-source geospatial proxies and satellite observations within a hybrid CNN-LSTM architecture to downscale and correct coarse global inventories (EDGAR/CAMS) to a 1 km resolution. The model integrates static urban features (nighttime lights, road density) and dynamic atmospheric data (Sentinel-5P trace gases, ERA5 meteorology) to produce spatially refined emissions for key pollutants including PM2.5, NOx, and black carbon. These emissions drive a coupled WRF-CMAQ chemical transport model, with subsequent systematic bias correction using an XGBoost algorithm trained on ground observations. Results reveal that global inventories systematically underestimate localized combustion sources in this urban environment. The downscaling increased black carbon emissions by ≈101.8%, primary PM2.5 and NOx by 66.8 and 71.1%, and aromatic VOCs by over 80%. Implementing the refined inventory in WRF-CMAQ significantly improved simulated surface concentrations: correlations with observed PM2.5 and NO2 increased by over 200% and 95%, respectively, during the dry season, while root mean square error was reduced by 25–35%. The refined model successfully resolves sharp urban emission hotspots aligned with transportation corridors and dense settlements. This work provides a validated, high-resolution dataset and a reproducible methodology that enhances exposure assessment and supports targeted air quality management in data-sparse regions of the Global South.