City-scale Modeling of Urban Heat Islands for Kolkata

Abstract

The Kolkata metropolitan area (KMA) is one of the sub-tropical urban environments experiencing swift and incessant urbanization process that has resulted in significant energetic differentiation in rural-urban domain as advection of heat wave and urban heat island (UHI) effect, it has to persuade on urban climate, biological environment and socio-economic atmosphere of urban society. The preparation of UHI susceptibility zonation is the preliminary measure for UHI risk assessment and hazard mitigation. The present study has been adopted the city-scale modeling of UHI by means of geographic information system (GIS) based statistical models for building the UHI susceptibility zonation using remote sensing (RS) data and other ancillary data. Initially, the UHI inventory map with 350 random pixels were extracted from mono-window algorithm (MWA) derived land surface temperature (LST) map using e-cognition approach. As such, UHI locations in LST map were then split into a ratio of 70/30 for building the UHI models and model validations. Finally, a spatial database of socio-economic, structural and radiative drivers with nine UHI conditioning factors has been prepared, including population distribution, land use and land cover (LULC), building material, building height, building roof type, building roof reflectance, building age, building association and road pavement and these database were extracted from multi-spectral scanning (MSS), thematic mapper-5 (TM-5) of Landsat images, Google Earth (GE) historical images and OpenStreetMap (OSM) with intensive rapid visual field survey (RVFS). The geo-spatial relationships between UHI inventory’s pixels locations and nine conditioning thematic factors were recognized by using four GIS-based statistical models i.e. analytical hierarchy process (AHP), two-class kernel logistic regression (KLR), support vectors machines (SVM) and spatial multi-criteria evaluation (SMCE) model. These models were constructed based on of training dataset and model-derived results have been validated and compared with the area under receiver operating characteristic (ROC) curve, kappa index and five different statistical evaluation measures to corroborate the noteworthy differences on the overall performance. The results of goodness of fit are of 86, 87, 85 and 89% and corresponding prediction capabilities are of 81, 85, 82 and 87% for AHP, KLR, SVM and SMCE models respectively. The statistical measures show that the SMCE model gives up overall better performance and precise results than the AHP, KLR and SVM models. The KLR and AHP models have produced to some extent better results than the SVM model in provisions of positive spatial prediction values. Hence, the study revealed that SMCE and KLR are the promising physical data mining approach to be considered to map the spatiality of UHI susceptibility zonation.