Multi-Source Data Fusion using Graph Attention Residual GCN for Urban Air Quality Prediction

¹ Department of Electronics and Communication Engineering, Nitte Meenakshi Institute of Technology, Nitte (Deemed to be University), Bengaluru, India
² School of Computer Science and Engineering, R V University, Bengaluru, India
³ Department of Electronics and Telecommunication Engineering, M S Ramaiah Institute of Technology, Bengaluru, India
⁴ Department of Electronics and Communication Engineering, Ballari Institute of Technology and Management, Ballari, India
⁵ Department of Computer Science and Engineering, Vignan's Foundation for Science, Technology and Research, A Deemed to be University, Hyderabad, India

^* Corresponding author: shivaprasad@msrit.edu

Abstract

Air quality forecasting has become essential for sustainable urban management and public health. However, traditional deep spatio-temporal methods often rely on Dynamic Time Warping (DTW), which introduces high computational costs and limited scalability, making real-time forecasting challenging. Therefore, lightweight framework combining Graph Attention Networks and Residual Graph Convolutional Networks (GAT-ResGCN) is proposed for efficient, fine-grained PM2.5. Initially, meteorological time-series data were gathered from India’s Air Quality Index dataset and satellite-derived PM2.5, maps from Modern-Era Retrospective Analysis for Research and Applications (MERRA-2) reanalysis dataset. Preprocessing included missing value imputation, Z-score normalization, and temporal alignment. Subsequently, GAT dynamically models spatial dependencies using k-nearest neighbor (k-NN) sparsification and Exponential Moving Average (EMA) updates. Furthermore, Spatio-temporal modelling was achieved through a Dynamic Spatio-Temporal Graph Convolutional Network (DSTGCN). Moreover, it integrates GAT-based graph convolutions, dilated Temporal Convolutional Networks (TCN) for long-range patterns, and temporal attention to highlight critical intervals. In addition, the Residual Network processes remote-sensing images to extract spatial pollutant patterns, which are fused with time-series features using a gated mechanism. The experimental results showed improved accuracy and scalability, achieving a root mean square error of 18.21, a Mean Absolute Error of 11.35, and an R² of 0.92.