A GeoAI framework for remote-sensing-driven traffic activity and emission intensity in a data-scarce megacity

Abstract

In rapidly expanding megacities across the Global South, a profound data scarcity paradox frequently prevents effective environmental monitoring and evidence driven urban planning. While severe traffic congestion generates acute localised pollution, the total absence of continuous municipal traffic counters, ubiquitous vehicle telemetry, and air quality monitoring networks leaves these environments empirically unmonitored. This study resolves this challenge by introducing a reproducible geocomputational framework designed to map microscale traffic activity and relative emission landscapes by fusing multisource geospatial data and deep learning workflows. Focusing on the unmonitored context of Kinshasa, Democratic Republic of the Congo, our methodology uses very high resolution 15 cm satellite imagery to train and deploy a localised YOLOv8 (You Only Look Once, version 8) computer vision model, establishing an empirical baseline of fleet distribution. These static counts are dynamically distributed across the urban road network using a probabilistic Hub and Spoke Gravity Model, which is weighted through a zero shot Large Language Model semantic reclassification of unstructured OpenStreetMap features. Reconstructed kinematic flows are subsequently processed through velocity dependent Tier 3 EMEP/EEA algorithms to capture the nonlinear emission penalties triggered by severe network gridlock. To ensure framework defensibility, an independent unmanned aerial vehicle observation campaign was executed across ten distinct urban sites. The drone data directly validate model estimated physical parameters, achieving robust correlations for traffic volumes (R2 = 0.826) and link speeds (R2 = 0.925). Because literature derived chemical emission factors cannot be validated in situ, all absolute outputs are subjected to a rigorous multiparameter sensitivity analysis to report findings strictly as bounded ranges. The results reveal a clear structural handover in the spatial activity profile of the city. Daytime pollution is heavily driven by the informal paratransit fleet, which accounts for a bounded range of 82.8% to 95.1% of total PM2.5 emissions along primary corridors despite representing a numerical minority of active vehicles. Conversely, nocturnal baseline emissions transition to a landscape dominated almost entirely by private passenger vehicles. All code, processing pipelines, and validation datasets are available in a FAIR compliant open access repository. This framework establishes a scalable, remote sensing driven blueprint for diagnosing traffic activity structures in unmonitored metropolitan regions worldwide.