Facility-scale carbon footprint retrieval using Sentinel-5P TROPOMI downscaling via a machine learning and data fusion approach
Abstract
This research addresses the limited regional applicability of Sentinel-5 Precursor (Sentinel-5P) TROPOMI, which provides daily global coverage of Methane (CH₄) and Carbon Monoxide (CO) due to its coarse spatial resolution. A downscaling methodology is proposed to estimate carbon footprint data at a finer spatial resolution (250×250 meters) by integrating data fusion, machine learning regression, and remote sensing techniques. Beyond TROPOMI-derived XCH₄ and XCO, the methodology incorporates additional data from multiple satellite platforms including MODIS and Sentinel-2, as well as ERA5 Reanalysis model outputs. These serve as inputs to four machine learning regressors: XGBoost, Random Forest (RF), Multilayer Perceptron (MLP), and Bayesian Ridge for each gas. Ground-truth measurements for training and validation were collected from 12 globally distributed TCCON stations to ensure robust generalization across diverse geographic conditions, vegetation cover thicknesses, land covers, and elevations. Among all tested models, XGBoost delivered the strongest performance. For XCH₄, it achieved an RMSE of 0.01 ppb, NRMSE of 5.92%, bias of -0.0011 ppb, MAE of 0.0073 ppb, and R² of 0.85. For XCO₂, the corresponding metrics were RMSE of 1.761 ppm, NRMSE of 5.86%, bias of -0.043 ppm, MAE of 1.26 ppm, and R² of 0.89. Qualitative validation was also conducted using the suspension of oil refinery operations in Grangemouth, Scotland in April 2025. The XGBoost models demonstrated strong correlations, capturing a pronounced decline in both CH₄ and CO₂ concentrations before (January 2023) and after (January 2026) the suspension.