ST-MFXAI: Research Methodology and Experimental Design for Spatio-Temporal Multi-Sensor Fusion with Explainable AI in Urban Air Quality Prediction

Research into urban air-quality forecasting suffers from three structural gaps: spatial and temporal models function in isolation instead of via deeply integrated learning; explainability is treated as an after-thought via post-hoc methods; and cross-city generalizability is seldom verified. To bridge these gaps, this paper introduces the methodology for ST-MFXAI (Spatio-Temporal Multi-sensor Fusion with explainable AI). The proposed framework natively embeds a Graph Convolutional spatial component, a Transformer temporal block, and a gated fusion mechanism to output four distinct explanations (SHAP, LRP, Grad-CAM, and attention profiles) directly. This study deploys a time-split pipeline across three historical testbeds: Beijing Multi-Site AQI, KDD Cup 2018, and OpenAQ. This research formalizes four testable hypotheses that directly target multi-modal accuracy, inter-station spatial sensitivity, XAI fidelity, and a strict 15% R² degradation cap during cross-city transfer rather than using standard testing. Operationalization of these core claims into measurable outcomes relies on a comprehensive four-tier validation protocol tracking ablation, state-of-the-art benchmarks, spatial transfer, and explanation accuracy.