Maximum Nighttime Urban Heat Island (UHI) Intensity Simulation by Integrating Remotely Sensed Data and Meteorological Observations

TLDR

Remote sensing of the urban heat island has mainly relied on simple correlations, with few studies examining its temporal evolution in relation to climatic and meteorological factors. The study aimed to assess the feasibility of support vector machine modeling for daily maximum nighttime UHI intensity in Beijing using MODIS land products and meteorological observations. A Gaussian surface model first calculated the MNUHII, after which SVM regression was trained on NDVI, surface albedo, aerosol optical depth, relative humidity, sunshine hour, and precipitation. SVM predictions were accurate to 0.8–1.3 °C and outperformed multiple linear regression and ANN, with RH and AOD as the most influential factors and prior precipitation significantly reducing the UHI intensity.

Abstract

Remote sensing of the urban heat island (UHI) effect has been conducted largely through simple correlation and regression between the UHI's spatial variations and surface characteristics. Few studies have examined the surface UHI from a temporal perspective and related it with climatic and meteorological factors. By selecting the city of Beijing, China, as the study area, the purpose of this research was to evaluate the applicability and feasibility of the support vector machine (SVM) technique to model the daily maximum nighttime UHI intensity (MNUHII) based on integration of MODIS land products and meteorological observations. First, a Gaussian surface model was used to calculate the city's MNUHIIs. Then, SVM regression models were developed to predict the MNUHII from the following variables: the normalized difference vegetation index (NDVI), surface albedo, atmospheric aerosol optical depth (AOD), relative humidity (RH), sunshine hour (SH), and precipitation (PREP). Results demonstrate that the accuracy of the SVM regression in predicting the MNUHII was around 0.8°C to 1.3°C; in addition, the SVM regression outperformed the multiple linear regression and the artificial neural network with backpropagation. A scenario analysis indicates that the relationships between the MNUHII and its influencing factors varied with time and season and were impacted by previous precipitation. The RH and AOD were the most important factors that influenced the MNUHII. In addition, previous precipitation could significantly mitigate the MNUHII. The results suggest that future investigations on the surface UHI effect should consider the climatic and meteorological conditions in addition to the surface characteristics.