Spatial-and-local-aware deep learning approach for Ground-Level NO2 estimation in England with multisource data from satellite-based observations and chemical transport models

• Efficient TROPOMI NO 2 imputation using a deep neural network. • Developed a spatial-and-local aware deep learning approach for NO 2 estimation. • Proposed model outperforms CNN and traditional machine learning models. • Model transparency enhanced with interpretable SHAP explanations. High-resolution near-surface NO 2 data are crucial for monitoring air pollution dynamics. Satellite-based machine learning models are commonly used to estimate NO 2 concentrations, but tailoring advanced deep learning techniques to specific environmental problems remains challenging. This study applies a two-stage deep learning approach to estimate ground-level NO 2 concentrations in England at a 1 km spatial resolution from 2019 to 2021. Initially, we imputed the TROPOMI NO 2 column density to a continuous 1 km resolution. We then developed an efficient spatial-and-local-aware deep learning network (SLNet) for NO 2 mapping by integrating the imputed TROPOMI NO 2 data with multi-source information from meteorology, chemical transport model (CTM) simulations, and other auxiliary predictors. To address the translation invariance of convolutional neural networks (CNNs), we combined a local channel to identify spatial heterogeneity in the model. Our imputed TROPOMI NO 2 surfaces, which initially covered only 34.12 % of valid data, achieved full coverage with reliability and continuity at 1 km spatial resolution. Cross-validation confirmed that the SLNet model outperformed other state-of-the-art methods in estimating ground-level NO 2 . The prediction model achieved R 2 values of 0.914, 0.919, and 0.887 for 2019, 2020, and 2021, respectively, and performed well in urban regions. Additionally, the Shapley Additive Explanations (SHAP) method revealed that features such as satellite and CTM NO 2 , precipitation, green space, and road density significantly contributed to estimations through both spatial and local channels. The mapping results closely aligned with ground-level observations and accurately captured spatial variations. This study advances NO 2 concentration estimation by applying adaptable deep learning techniques and enhances the understanding of air pollution dynamics.