Efficient Vision Transformer Framework: Near-Surface O₃ Prediction with Missing Satellite Data

This paper presents an efficient Vision Transformer (ViT) framework for predicting near-surface Ozone (O3) concentrations using incomplete satellite data. Industrial activities have significantly increased atmospheric O3, a pollutant with severe health and environmental impacts. Traditional ground observation techniques lack the geographical coverage provided by satellite data, but these satellite observations often suffer from continuity issues. To address these gaps, we propose a novel framework that integrates satellite and corresponding ground observation grid cells data to improve the accuracy of O3 predictions over large areas. Our primary contributions include: introducing a ViT framework that leverages multi-source data for accurate O3 prediction; using satellite data for training and ground observations for validation to reconstruct missing data and provide continuous O3 predictions. We demonstrate through experimental analysis that our framework achieves Root Mean Square Error (RMSE) values of 16.2 µg/m³, significantly outperforming four state-of-the-art models in prediction accuracy. The code for this research is available here: https://github.com/Prasanjit-Dey/O3_Prediction.