Mining and modeling the direct and indirect causalities among factors affecting the Urban Heat Island severity using structural machine learned Bayesian networks

Abstract

Urbanization, population growth, and climate change have several impacts on the environment including the extreme increase in temperature in urban areas, which is also known as the Urban Heat Island (UHI) effect. This paper presents a novel white-box data-driven structural learning Bayesian network model that (1) discovers knowledge from the data by identifying the key factors impacting the UHI severity; (2) captures the causal (direct and indirect) relationships between the different variables that influence UHI severity, and (3) represents the learned relationships into graphical networks that are both machine- and human-interpretable. Different Bayesian networks were developed based on a dataset comprised of 31 meteorological, socio-demographic, geographic, and land use/land cover factors gathered for the State of New Jersey, USA. Furthermore, the different Bayesian networks were assessed and compared to determine the optimal structure. Finally, the best model was validated on an unseen testing sample where an overall accuracy of 88.51% was obtained. The proposed optimal Bayesian network model was able to discover knowledge about 13 causal relationships between 12 variables (one of which is the UHI severity). The outcomes of this research are crucial for urban management and for proposing proper adaptation plans for the UHI effect.