Language Model Empowered Cross-Scale Framework for Urban Air Pollution Monitoring, Diffusion Analysis and Early Warning

Abstract

Modern urban morphology and local meteorological conditions create significant variations in pollutant dispersion across scales. Under global climate change and rapid urbanization, air pollution threatens public health and sustainable development, yet traditional monitoring fails to address both micro-level protection and macro-level governance—especially within the “street canyon–residential area” microenvironment where daily human activity concentrates. To bridge this gap, this study first develops an LLM-enhanced CFD system with atmospheric stability metrics to model pollutant dispersion from street canyons to residential areas. Since micro-level research cannot cover large regions, we further propose a context-aware human-machine collaborative drone framework integrating multimodal sensing, XGBoost prediction, and natural language interaction for autonomous surveillance, source localization, and short-term alerts. Results show that under stable conditions, peak CO in street canyons exceeds 80 µg/m³—over three times higher than under unstable conditions (5–35µg/m³)—with high-concentration plumes traveling long distances to downstream residential areas. The drone system achieves 0.34 m localization accuracy, a 1‑hour CO prediction RMSE of 2.3 µg/m³, 98% natural language interpretation accuracy, and sub‑3‑minute response to sudden pollution events. This cross-scale framework unifies micro-mechanisms and macro-monitoring, offering a practical solution for urban air quality management and healthy city construction.