Improved estimation of carbon dioxide and methane using machine learning with satellite observations over the Arabian Peninsula.

Estimating spatiotemporal maps of greenhouse gases (GHGs) is important for understanding climate change and developing mitigation strategies. However, current methods face challenges, including the coarse resolution of numerical models, and gaps in satellite data, making it essential to improve the spatiotemporal estimation of GHGs. This study aims to develop an advanced technique to produce high-fidelity (1 km) maps of CO and CH over the Arabian Peninsula, a highly vulnerable region to climate change. Using XGBoost, columnar carbon dioxide (XCO) and methane (XCH) concentrations using satellite data from OCO-2 and Sentinel-5P (the target variables) were downscaled, with ancillary data from CarbonTracker, MODIS Terra, and ERA-5 (the input variables). The model is trained and validated against these datasets, achieving high performance for XCO (R = 0.98, RMSE = 0.58 ppm) and moderate accuracy for XCH (R = 0.63, RMSE = 13.26 ppb). Seasonal cycles and long-term trends were identified, with higher concentrations observed in summer, and emission hotspots in urban and industrial areas. Comparisons with the EDGAR inventory highlighted the significant contributions of the power, oil, and transportation sectors to GHG emissions. These results demonstrate the value of high-resolution data for local-scale monitoring, supporting targeted mitigation strategies and sustainable policymaking in the region. Future work could integrate ground-based observations to further enhance GHG monitoring accuracy.