AI-Driven Urban Morphology: Mapping Pollution Exposure through Deep Spatial Learning and Transportation Flow Networks

Abstract

An AI-driven approach to urban morphology grounds pollutant-exposure rendering that harnesses deep spatial learning and the city’s transportation flow networks. Are pollutant-concentration maps a function of urban form? A twelve-layer deep residual learning model, trained for ten-hundred epochs and informed by a heterogeneous mix of spatio-temporal features, generates three baselines: full set, saturation point, and without vegetation. The proposed model adds flow-to-market and flow-to-source networks, and uncertainty evaluations incorporate the strengthening effect of vegetation alone. Pollution distribution in Los Angeles suggests a west-east dipole, with lower values in the city-centre, proximity to the ocean, and higher elevation.; contours of traffic-related pollutants expose a plurality of high-risk regions beyond the immediate vicinity of main roads. Does urban form matter? Qualitative–quantitative comparisons to baseline models demonstrate the effect: city shape factor, road density, functional diversity, and urban heat island are significant predictors. Theory and evidence could guide planners and decision-makers towards a model-driven design of urban forms promoting lower pollution exposure.

Urban morphology links city shape, form, and functional distribution to pollutant concentrations and the spatial–temporal continuum of perceived exposure. The transport–land-use framework that links transportation-related emissions to their sinks and sources motivates a closed-loop design of spatial concentration maps via an urban-flow proxy-consumer approach. Deep spatial-learning techniques evade the need for extensive knowledge of physical processes by automatically learning spatial dependencies, while data fusion expands the range and type of such dependencies