Machine learning estimates for G20 subnational GHG emissions from 2000-2020 using self-reported emissions data

Reliable, comparable greenhouse gas (GHG) emissions data at the subnational level remain scarce, despite growing expectations for cities and regions to lead on climate action. Inconsistent reporting, methodological variation, and limited coverage of self-reported inventories hinder efforts to track progress and guide mitigation opportunities. To address these challenges, we develop a machine learning (ML) framework to estimate annual Scope 1 and 2 CO2-equivalent emissions for subnational jurisdictions in G20 countries from 2000 to 2020. Our approach integrates publicly available geospatial, socioeconomic, and environmental data with self-reported inventories where available, and aligns predictions with subnational administrative boundaries. Compared to traditional downscaling or proxy-based approaches, our model improves spatial relevance and predictive performance while capturing locally specific emission drivers. This globally consistent, administratively-aligned dataset can serve as a baseline for assessing climate progress, especially in data-poor or inconsistent reporting contexts, and supports more targeted, data-informed policy decisions for urban and regional decarbonization.