Top-Down Mapping of CO₂Emissions: A Finer-Scale Approach with Remote Sensing and AI

Elevated atmospheric carbon dioxide (CO2) levels contribute to global warming, necessitating urgent emission reduction. Identifying CO2 sources is crucial. This study develops end-to-end models for high-resolution national CO2 estimation using remote sensing. Our methodology involves three steps. First, a machine learning-based model establishes relationships between satellite-derived column average CO2 (XCO2) and weather conditions, including anthropogenic proxies. This model generates daily 1 km2 spatial XCO2 maps. The second step separates dominant accumulated XCO2 (XCO2bg) and regional enhancement (ΔXCO2) due to anthropogenic activities, challenging due to ΔXCO2 being small (ΔXCO2 ≫ XCO2bg) and often near measurement noise. Addressing this, we adopt a geometrically connected segmentation to identify emission and non-emission sources, establishing XCO2-NO2 relationships for ΔXCO2 maps at a weekly frequency. The final step involves ΔXCO2 to CO2 emission conversion, challenging due to dispersion processes. We customize an integrated mass balance method for weekly, 1 km2 spatial XCO2 emissions mapping. Our approach aligns closely with reported annual the Kingdom of Saudi Arabia (KSA) emissions, showcasing high-resolution emissions tracking and a departure from traditional bottom-up approaches, enabling near real-time (NRT) finer to country-level emission monitoring, circumventing delays associated with annual reporting.