Air quality degradation is a major environmental and health concern with traditional monitoring systems still being expensive and challenging to set-up in geographically diverse areas. In this paper, Vision Transformer (ViT), Graph Convolutional Network (GCN) and Graph Attention Network (GAT) are combined with a vision-driven deep learning model named AtmosViT-GAGNet that can be used for atmospheric quality assessment and real-time Air Quality Index (AQI) estimation. The framework is used to extract visual characteristics from images acquired in the atmosphere, model the relationship between visual and environmental characteristics and infer influential characteristics of the atmosphere that can be driven by attention based learning. The fused representations aid AQI regression and five-level classifications of the atmospheric conditions. Evaluation based on the benchmark with more than 12200 images of atmosphere with labels gained the result of 98.84% classification accuracy with 0.91 F1-Score in relation to fog density, mean absolute error in relation to AQI of 8.3 and R² equal to 0.87. The results highlight the potential of combining visual intelligence with graph learning for use in scalable systems for atmospheric monitoring.

Atmospheric Quality Assessment, Air Quality Index Prediction, Vision Transformer, Graph Convolutional Network, Graph Attention Network, Deep Learning, Environmental Monitoring.