Deep Learning-Enabled Spatiotemporal Monitoring of Global Air Pollutants Using Remote Sensing: Insights into Data-Scarce Regions

Nitrogen dioxide (NO₂) and Sulfur dioxide (SO₂) are important targets for monitoring atmospheric quality. Accurate ground concentration measurements are fundamental steps in pollution prevention and risk reduction. The scenario poses significant challenges for air quality monitoring in arid environments, particularly in the Middle East and North Africa (MENA) region, due to rapid urbanization and the scarcity of ground-based sensor networks. While satellite remote sensing, such as the Sentinel-5P TROPOMI mission, provides synoptic global coverage, its usefulness for assessing public health is limited by the difference between column densities and surface-level concentrations. This paper presents a novel hybrid AI framework that combines spatiotemporal inversion with deep learning-based forecasting to address this gap, particularly in ground data-scarce regions. Our approach follows a thorough three-phase framework. First, we created the Dynamic Urban-Met Integration (DUMI) database. This cohesive spatiotemporal tensor integrates trace gas data from Sentinel-5P/TROPOMI (NO₂, SO₂), MERRA-2 meteorological reanalysis data, and urban growth statistics from the UN World Urbanization Prospects (WUP) 2025. To overcome the resolution difference between satellite (~5.5 km) and meteorological (~50 km) data, we employed a zonal spatial aggregation algorithm, implemented within the Google Earth Engine (GEE), to synchronize multi-resolution sources within a standardized 30 km urban airshed for 100 global cities spanning from 2019 - 2025. Second, we employed a Homogeneous Domain Adaptation approach to address the challenge of insufficient local ground-truth data. In particular, we trained an Extreme Gradient Boosting (XGBoost) regressor using data from a "Source Domain" comprising 20 data-rich U.S. cities, selected as climatic analogs with urban typologies similar to data-scarce regions, including industrial congestion, traffic patterns, desert dynamics, and other urban features. This method facilitated the approximation of the nonlinear physical transfer function (C_surf = f(N_col, PBLH, Wind)), which is influenced by wind dynamics and the Planetary Boundary Layer Height (PBLH). Lastly, we used a 12-month sliding window to train a stacked deep learning forecasting model, such as a Long Short-Term Memory (LSTM) network, using the rebuilt "Synthetic History." With this configuration, the model can anticipate future trajectories under the urban growth scenarios of those cities from 2026 – 2030 and incorporate seasonal volatility. Preliminary validation against held-out US EPA ground station measurements (2019-2025) shows that the inversion model successfully captures the physics of atmosphere dilution, with (R²) values of 0.998 for NO₂ and 0.992 for SO₂ using monthly mean data. SHAP (SHapley Additive exPlanations) analysis provides additional evidence of the model's physical consistency by revealing that the AI autonomously learned the strong inverse relationship between PBLH and surface concentrations (the "Lid Effect"), validating its transferability to new regions. Preliminary testing in Los Angeles and Seoul indicates that the LSTM can sufficiently generalize to predict seasonal volatility and pollution spikes, with an (R²) value of 0.84 & 0.82, respectively. This approach provides a scalable "Virtual Station" infrastructure that gives policymakers a quantitative tool to assess the environmental effects of rapid urbanization in data-poor dry regions.

Keywords: GeoAI, Nitrogen dioxide (NO₂) & Sulfur dioxide (SO₂), Inversion, Remote Sensing, XGBoost, Sentinel-5P, Deep Learning, LSTM, SHAP, Saudi Arabia.