Predicting Urban Heat Island severity on the census-tract level using Bayesian networks

Abstract

Urban development and population growth have resulted in several critical impacts on the society, environment, and economy. One of the main impacts is the increase in temperature observed in urban areas, also known as the Urban Heat Island (UHI) effect. UHI has become the focus of several research studies due to the associated negative implications it causes. Despite that some research efforts examined different characteristics of the UHI phenomenon, there is a gap in the literature in terms of developing interpretable machine learning models that can accurately estimate or predict the severity of UHI (rather than air temperature or other UHI characteristics) as well as that can provide predictions on the census-tract level in the US. To that extent, this paper fills this gap by developing a knowledge-based white-box Bayesian network model that predicts UHI severity based on demographic, meteorological, and land use/land cover factors. First, a dataset for all census tracts in the state of New Jersey, USA was developed, which is comprised of 13 independent variables or factors affecting UHI severity. Second, expert knowledge was obtained from 10 UHI experts using the systematic three-round Delphi method to develop four different Bayesian network models. Third, the performance of the four Bayesian models was assessed and compared to choose the optimal model with the highest accuracy. Finally, sensitivity analysis was conducted to assess the influence of each key factor on the UHI severity. The results showed that the optimal model is a tree-augmented Bayesian network that can predict or estimate the UHI severity with an accuracy of 87.88%. The outcomes of this paper also reflected that the following 8 variables are the key factors that impact UHI severity: NDVI during winter season, NDVI during summer season, imperviousness, tree canopy, building area, population density, water body areas, and annual rainfall. The findings also identified the land use/land cover category as the major category affecting UHI severity compared to demographic- and meteorological-related factors. The proposed white-box Bayesian network model in this paper adds to the body of knowledge by allowing practitioners and researchers to perform micro-level UHI predictions and inferences about future or unknown UHI severity pertaining to individual communities or small/localized geographical areas. This enables more focused and targeted UHI mitigation actions to be planned, designed, and implemented in the most affected communities, which ultimately results in better decisions, actions, and outcomes to reduce or address the UHI effect.