Answers are grounded in this paper and cite their sources.
Demonstrates improved nitrogen dioxide estimates in urban environments, indicating effective integration of satellite and ground data.
Satellite-derived NO₂ observations play an important role in monitoring atmospheric pollution; however, their utility at local scales is often constrained by coarse spatial resolution, data gaps caused by cloud cover, and limited representation of surface-level variability. To address these challenges, this study presents a machine learning–driven pipeline for reconstructing higher-resolution surface NO₂ concentrations from satellite observations by integrating remote sensing data with ground-based measurements and auxiliary geospatial variables. The proposed framework employs Long Short-Term Memory (LSTM)–based temporal imputation to address missing satellite observations and applies spatial interpolation to improve data density prior to modeling. Multiple regression strategies were evaluated, including satellite-to-ground, ground-to-satellite, and satellite-to-satellite mappings, using learners such as Random Forest, XGBoost, LightGBM, and neural networks. Based on comparative evaluation using standard regression metrics on held-out validation data within the study region, Random Forest regression was selected for subsequent analysis. The final configuration combines machine learning–based prediction with spatial interpolation and ancillary inputs, including vegetation indices and land cover, to enhance spatial resolution from approximately 18,000 to 56,000 grid points over the study area. In addition, a web-based visualization platform is developed to enable interactive exploration of the reconstructed NO₂ fields and facilitate user-oriented dissemination of air quality information. The results indicate the feasibility of integrating satellite data, ground observations, and auxiliary features to generate fine scale NO₂ estimates within the temporal and geographic scope of this study. Temporal resampling validation across sequential sub-periods confirms stable model performance over the study window, and out-of-bag uncertainty estimates provide per-grid-cell confidence bounds, strengthening the reliability and interpretability of the reconstructed NO₂ fields.
No takes yet. Share an insight, caveat, or question.
Rangavajjula et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: