*4.4. Various Statistical Models*

Statistical models and machine learning have also been proposed to measure UHI [227]. Among these studies, a Gaussian surface model has been utilized the most because it can provide not only the intensity, but also the spatial extent and the central location of the UHI. The kernel convolution method has also been proposed to study UHI effects because of its high efficiency in characterizing the temperature values over space in a continuous surface [227]. Chun and Guldmann [77] explored the urban determinants of UHI using two-dimensional (2D) and three-dimensional (3D) urban information as the input for spatial statistical models. The results show that solar radiations, open spaces, vegetation, building roof-top areas, and water strongly impact surface temperatures, and that spatial regressions are necessary in order to capture the neighboring effects. Recently, Li et al. [81] estimated UHI intensity by linear regression functions between LST and regionalized ISA. These statistical models could avoid the bias caused by the definitions of urban −rural areas or the choice of the representative pixels, and thus facilitate the comparison of UHI among cities. Szymanowski and Kryza [228] addressed the issue of the potential usefulness of remotely sensed data and their derivatives for UHI modeling. Statistically significant models explained 71% to 85% of the air temperature variance. It has been stressed that remotely sensed data are important sources to model urban air temperature heat islands. However, in all of these studies, such models worked less effectively in cities frequently covered by clouds, in arid landscapes, and in urban agglomerations, so they have only been applied in a few UHI studies to date. Recently, Lai et al. [79] published the statistical estimation of next-day nighttime surface urban heat islands of selected cities. Most previous studies modelled the SUHI variations for the past period, but rarely investigated the estimation for future UHIs, especially at the daily (i.e., day-to-day) scale. To address this issue, this study incorporated both meteorological and surface controls to estimate next-day nighttime UHIs using a support vector machine regression (SVR) model. Some uncertainties exist in terms of the Gaussian modelling and UHI estimators, which may limit estimation accuracy. Nevertheless, by providing a feasible ye<sup>t</sup> simple approach for estimating next-day nighttime UHIs, this study fills a knowledge gap in the UHI estimation and is helpful for supporting adaptation to and mitigation of UHI and UHIRIP.
