Satellite image processing for mapping air pollution dispersion near biorefinery sites

Abstract This study presents a satellite‐based analytical framework that integrates Nitrogen Dioxide (NO 2 ) observations from Sentinel‐5P/Tropospheric Monitoring Instrument (TROPOMI) with meteorological variables (temperature and humidity) from Moderate Resolution Imaging Spectroradiometer (MODIS) and wind speed data from Gridded Surface Meteorological Dataset (GRIDMET) to assess air pollution dispersion near bio refinery sites. The framework uses Long Short‐Term Memory (LSTM) neural networks enhanced with Time2Vec temporal encoding to forecast NO 2 concentrations across three Saudi Arabian cities—Al Riyadh, Al Jubail, and Najran—using daily measurements collected between August 2018 and April 2023. Data preprocessing includes cloud masking through Google Earth Engine (GEE) quality‐control filters, kriging‐based interpolation for meteorological variable alignment, and min–max normalization. The Time2Vec‐LSTM model achieves a mean absolute error of 41.57 μg/m 3 for Al Riyadh, with a Root Mean Square Error (RMSE) of 54.44 μg/m 3 and a Mean Absolute Percentage Error (MAPE) of 33.07%, representing a 2.8% improvement over the baseline LSTM. Spatiotemporal analysis indicates distinct seasonal variations, with peak NO 2 concentrations during winter (attributed to increased heating emissions) and minimum levels during summer (due to enhanced atmospheric dispersion). Overall, the proposed framework addresses monitoring gaps in regions with limited ground‐based air quality networks and offers valuable decision‐support capabilities for environmental policy development and sustainable urban planning.