Title Page Coclustering-enhanced Space-Time Light Gradient Boosting Machine for Mapping Long-term Ambient NO₂ and Associated Health Effects in China from Satellite observations

Chronic exposure to Nitrogen dioxide (NO₂) is strongly associated with adverse health effects. Integrating satellite observations with in-situ measurements enables spatially continuous assessment of ambient NO₂ concentrations and their health effects. However, existing machine learning-based models for remote sensing of long-term ambient NO₂ in China may not fully consider heterogeneity existing in NO₂ concentrations and their association with controlling factors, resulting in low accuracy. To address the issue, we developed an innovative hierarchical NO₂ remote sensing framework that integrates the co-clustering algorithm with space-time tree-based method, named Coclustering-enhanced Space-Time Light Gradient Boosting Machine (Coclust-STLightGBM). To consider regional heterogeneity of environmental and anthropogenic controls on NO₂, the study area was first delineated into sub-regions using the coclustering method. Then region-specific parameterized space-time LightGBM (STLightGBM) models with explicit spatial and temporal features incorporated were trained for remote sensing of ambient NO₂, to consider local variations of NO₂ over space and time. Experiments showed that Coclust-STLightGBM significantly outperformed STLightGBM and LightGBM, with out-of-sample (out-of-station) R² increasing by 0.06 (0.07) and 0.13 (0.11) while RMSE reducing by 1.47 (1.26) μg/m³ and 2.60 (1.75) μg/m³ in China. Results revealed pronounced spatio-temporal heterogeneity in ambient NO₂ and associated mortality burden in China during 2005-2020: high NO₂ in central and eastern China during winter, and low levels in the west during summer. Concentrations rose from 2005 to 2014, driven by industrial growth, then declined through 2020 following the “Ten Measures for Air Pollution Control”. The COVID-19 pandemic in 2020 further reduced NO₂ by up to 11.23% nationally and 45.25% in Hubei. Attributable mortality patterns closely mirrored NO₂ distributions. These findings demonstrate that Coclust-STLightGBM is a powerful tool for accurately mapping remote sensing of long-term ambient NO₂ concentrations and associated health impacts in China.