AI-Assisted Monitoring of Atmospheric Greenhouse Gases Over Bangalore Using Satellite And Reanalysis Data

Abstract

Cities concentrate greenhouse gas emissions but rarely have the dense ground monitoring needed to locate them. This study tested whether satellite observations and machine learning can close part of that gap over Bangalore, India. Monthly Sentinel-5P columns of methane, carbon monoxide and nitrogen dioxide for 2020 to 2024 were combined with ERA5 reanalysis on a 0.05° grid, giving 5,619 cell-month records. Random Forest, XGBoost and LightGBM were trained to predict methane from 22 predictors and evaluated under shuffled, year-blocked and cell-blocked cross-validation. Hotspots were detected with HDBSCAN and the models interpreted with SHAP. All three algorithms agreed to within 0.0007 in R², reaching 0.956 on a shuffled hold-out split. That figure did not survive stricter validation. Blocking by grid cell left it at 0.936, while blocking by year drove every model to an R² near −1.57, worse than predicting the training mean. SHAP explains the asymmetry: season and year carry 54.6 percent of the attributed importance, and methane varies by only 0.86 percent across the grid against a seasonal swing of about 62 ppb. Two nitrogen dioxide hotspots covering 16 of 100 cells were identified, both attributed to traffic and both over the densest part of the road network in independent land-use data. Elevated cell-months were flagged at 88.2 percent with a false positive rate of 2.8 percent. The framework supports spatial gap-filling, hotspot localisation and the ranking of elevated periods. It does not support forecasting. A single shuffled-split score would have hidden that.