Predictive Modeling of Urban Commute Emissions Using Hybrid Deep Learning and Geo-Spatial Data

Abstract

Predictive modeling of urban commute emissions through a hybrid deep-learning framework using geo-spatial data brings together four ingredients—imperfect past, uncertain future, neural wisdom, and classical science—designed to step in closer harmony with the breath of the streets. The problem is relevant and structured enough to spur each of the core components into play—the presence of explanatory variables suggests a traditional model, tempting a neural network at the same time and stimulating an interest in exploring the spatio-temporal aspect more deeply through a graph neural network, grid-based local encoding with global context, attention, and time series ingredients. The results connect these modelling pieces yet the exploration of meanings in the predictions and the data-shaping and curation effort are much richer thanks to the modelling soup. Three case studies illustrate what emissions narratives say at different scales and how the models help in their interpretation: the heart of the financial district women’s warning and the ship transports use too much of Manhattan’s air towards the river for three critical hours every day; the Riverside Corridor at Dawn, still and silent yet nevertheless working and alive, with mostly pies, tubes and the dependence on the classic modes for support; crossing the DCP, transit, driving and that drive-share always empty at 7 am, but then goes up and gets filled on the drop in.

Urban communities continue to grow, putting pressure on infrastructure systems, services and the environment. Deleterious impacts from transport mostly arise in urban agglomerations, sources of large quantities of greenhouse gases and other pollutants, that being time-lice and place-limited – more than 66% of urban citizens breathe the air classified as being of poor quality by WHO during at least one month in a year. Investment decision making and implementation, in urban or peri-urban areas, require reliable quantitative estimates of traffic emissions that include the actual situation and that of future scenarios to support the so-often neglected what-if issues.