100 m PM2.5 mapping from SDGSAT-1 TOA reflectance: Model development and evaluation

Satellite-based fine-resolution (∼100 m) PM 2.5 mapping remains challenging because broadband multispectral imagery struggles to decouple land and atmospheric signals, limiting accurate local emission source detection in regions with sparse ground monitoring networks. Here, we introduce LAD-GAT, a novel deep learning framework for 100 m-resolution PM 2.5 estimation from SDGSAT-1––the first science satellite mission dedicated to the UN Sustainable Development Goals. Specifically, LAD-GAT builds a high-dimensional scene-attribute graph by combining PM 2.5 -relevant geographical features, meteorological dynamics, top-of-the-atmosphere (TOA) reflectance, and estimated surface reflectance (SR). A specialized land-atmosphere decoupling (LAD) module is introduced to separate latent aerosol signals from ground surface contributions, and a graph attention network (GAT) models nonlinear associations between in-situ PM 2.5 observations and the input graph structure. In 10-fold cross-validation, LAD-GAT achieved RMSE = 5.042 μg m −3 (R 2 = 0.875) using SDGSAT-1 TOA reflectance, and RMSE = 9.428 μg m −3 (R 2 = 0.862) with Sentinel-2 TOA reflectance. Incorporating daily SR yielded an 8.68 % accuracy gain over TOA reflectance alone and outperformed multi-day composites by 7.27 %, highlighting the benefit of accounting for SR dynamics in fine-scale PM 2.5 mapping. Overall, leveraging the proposed novel LAD-GAT method, SDGSAT-derived PM 2.5 estimates rival those from Sentinel-2, providing fine-scale data to better support SDG 11.6.2 monitoring and targeted air-quality interventions. • Developed a novel LAD-GAT model for fine-scale PM 2.5 mapping using SDGSAT-1 data. • Synthesized daily SDGSAT-like surface reflectance by fusing Sentinel-2 and MODIS data. • Demonstrated that SDGSAT-1 performance in PM 2.5 mapping is on par with Sentinel-2. • Achieved a PM 2.5 estimation RMSE of 5.042 μg m −3 from SDGSAT-1 TOA reflectance. • Enhanced PM 2.5 modeling accuracy by 8.68 % through incorporation of surface information.