Investigating important urban characteristics in the formation of urban heat islands: a machine learning approach

As UHIs occur as a result of the land cover transformation, primarily replacement of natural vegetation and natural land cover by impervious surfaced associated with urban land uses, previous empirical UHIs research has mainly focused on the causal relationship between remotely sensed land surface temperature and remotely sensed physical land surface characteristics of urban areas such as elevation and slope [15, 16], land cover types [5, 10, 17, 18], percent impervious surface [5, 15, 19], percent tree canopy cover [5, 7, 19], vegetation abundance indices [16, 18‐20].

There is another stream of UHIs research that investigated social vulnerability to extreme heat-related weather events. Social vulnerability to environmental hazard means the relative potential for physical harm and social disruption to subpopulations of societies and their larger subsystems based on socioeconomic status, age, gender, race and ethnicity, family structure, residential location and other demographic variables [21]. In the studies investigated the relationship between extreme heat relate weather events and socioeconomic status of the area, socioeconomic variables including total population [12], population density [15, 16, 19], total housing units [12], gender [12], race and ethnicity [12, 14, 16, 22, 23], number of older and younger population [12, 14, 23, 24], education [14, 23], median house value [12, 16, 25], median household income [13, 22], poverty [14, 22‐24], housing tenure [22], number of mobile homes [12] were most widely used and found significant.

Starting from the above-stated discussion of what biophysical and socioeconomic variables are to be included in the analysis, the unit of analysis, which means in what geographical unit all of the variables to be integrated is another important issue to be discussed. As Stone and Norman [7] addressed earlier in their study, in order to associate aforementioned remotely sensed urban surface attributes and socioeconomic characteristics with the planning and planning decision-making, it is essential that these attributes should integrate into a planning-relevant scale.

However, there is no commonly agreed upon guidelines for aggregating data into the certain unit of analysis or theory that provides any arguments in favor of certain data integration method. So many empirical studies have selected unit of analysis using cumulative results from prior empirical studies. There are studies that have adopted a unit of analysis consistent with the resolution of a satellite image, a grid cell [5, 18, 19, 25], and others used a unit that of socioeconomic variables were measured such as County [13], Census tract [16, 17], and Census block group [14, 26]. To the best of my knowledge, Stone and Norman [7] is the only study that used a parcel as a unit of analysis.

In order to associate aforementioned research results with the practice of urban planning and policy, it is essential to address the relationship between land surface temperatures in planning relevant urban scale. If urban planners may be able to mitigate the formation of UHIs through the modification of specific zoning and various regulation, urban physical characteristics highly correlated with higher land surface temperature in planning relevant urban scale should be identified first. As socioeconomically vulnerable population should be identified together with physical characteristics of the urban area, socioeconomic characteristics highly associated with higher land surface temperature also need to be addressed in that relevant urban scale.

However, it is methodologically not so simple because remotely sensed surface temperatures are measured in micro scale but resulting socioeconomic vulnerabilities are presented in much larger scale. Also, to effectively integrate the findings of the UHIs research, the result should be addressed in planning relevant scale. However, neither micro scales such as a pixel nor macro scales such as a Census block group and county cannot be directly used by urban planners to reduce the magnitude of UHI effects. This study tried to overcome this discrepancy in scale issue in following two ways: first, divide dataset by planning zones, as different planning zones represent different human activities so will need different UHIs mitigation strategies, second, integrate urban biophysical characteristics which measured in microscale and socioeconomic characteristics which measured in the Census block group level into land parcels for parcels are the spatial unit of urban planning and design.

To investigate what are the most important determinant factors in the formation of the UHIs, who are the most severely affected by the consequence of the UHIs phenomenon, and where the phenomenon is the most intense, this study applied a machine learning approach. As there is no theory that explains the aforementioned relationship, thus the machine learning technique, which allows the computer to learn without being programmed and to identify patterns in data, was used to find highly determinant variables to the formation of UHIs in Marion County, Indiana. To quantify the causal relationship between the land surface temperature and explanatory variables, previous empirical studies have been used regression approaches such as simple regression and spatial regression approaches [19, 20, 26] and correlation analysis [18]. The PCA method, a mathematical procedure that transforms a number of correlated variables into a smaller number of uncorrelated variables called principal component, has been widely used in many of empirical UHIs and social vulnerability studies [12, 13, 19] to reduce the data dimension.

To the best of my knowledge, works of Hart and Sailor [5] and Rhee, Park, and Lu [20] are one of the few research that applied machine learning technique for the identification of highly determinant variable to the formation of UHIs. Hart and Sailor [5] applied Cubist method, a rule-based regression tree approach, to determine the most important variables affecting the UHIs intensity in Portland, Oregon. Rhee and others [20] applied the RF method to investigate the relationship between land cover pattern and surface temperature in Denver, Colorado to identify the relative importance of variables. Because there is no commonly accepted theory to guide modeling the relationship between urban attributes and UHIs, this study strongly believed the machine learning approach that allows the computers to learn from data without programming can be used as a powerful tool to identify the importance of urban attributes to the formation of UHIs. This study applies RF method to identify the relative importance of urban biophysical and socioeconomic variables to the formation of UHIs at the parcel level.

Marion County, Indiana is the largest county in the state of Indiana and the 55th largest county in the US by its population [27]. The county seat is Indianapolis, the state capital and the largest city in Indiana. Indianapolis is 15th largest US city by its population, and it is one of those US cities where UHIs have grown in largest numbers in recent years [2, 27] (Fig. 1).

A Landsat 7 Enhanced Thematic Mapper Plus (ETM+) image of Marion County, Indiana, path 21 row 32, which was acquired on 12 July 2013 was used for this study. During summer months between the year 2011 and year 2016, an image dated 12 July 2013 was chosen because of the higher atmospheric temperature before and after the image acquisition date, and also due to the lower cloud cover. Image used in this study was taken at EST 4:18 p.m. According to US Environmental Protection Agency (US EPA) and US National Oceanic and Atmospheric Administration (US NOAA), year 2006 through 2015 was the warmest decade on record since thermometer-based observations began from 1850, and year 2013 was one of the hottest years on record [3, 28]. The image was rectified to a common Universal Transverse Mercator coordinate system based on 1:24,000 scale topographic maps and was resampled to a pixel size of 30 m by 30 m using a nearest neighbor resampling algorithm. The root mean square error (RMSE) achieved during the rectification was less than 0.5 pixel.

The thermal infrared (TIR) band of an image was converted to surface temperature following the process suggested by Weng and others [18]. The digital number (DN) of Landsat 7 ETM+ TIR band was first converted into spectral radiance using Eq. (1), and then converted to blackbody temperature (T_B) under the assumption of uniform emissivity using Eq. (2) below: where L _λ is spectral radiance, T _B is effective at-satellite temperature, or blackbody temperature in Kelvin, K₁ and K₂ are pre-launched calibration constant [28].

For Landsat 7 ETM+, K₁ = 666.09 mW cm⁻² sr⁻¹ µm⁻¹, and K₂ = 1282.71 K were used [28]. The emissivity corrected land surface temperature (T_s) were finally computed as follows:where λ is the wavelength of emitted radiance, and here λ = 11.5 µm was used, and ε is spectral emissivity [29]. ENVI 5.3 software was used for the land surface temperature calculation. Figure 2 shows the spatial distribution of summertime land surface temperature of the study area.

The formation of UHIs is the result of a complicated combination of land cover changes, structural characteristics of the built area, meteorological conditions, landscape structure, and socioeconomic conditions of the urban area. The consequences of the UHIs phenomenon differentiate among various socioeconomic groups. Based on the in-depth literature review and the availability of data for the study area, this study selected following urban physical characteristics and socioeconomic characteristics to investigate the causal relationship with land surface temperatures (Table 1).

All of the remotely sensed data including land surface temperature and NDVI which derived from Landsat 7 ETM + and NLCD 2011 data including percent developed imperviousness and percent tree cover were measured in 30-m by 30-m resolution. Data explaining urban physical structure including average and maximum building height, and building footprint were calculated for each building using LiDAR image in 1-m by 1-m resolution using ENVI 5.3 software. All of the socioeconomic variables were collected at the level of Census block group from the year 2013 American Community Survey (ACS) to be consistent with the Landsat 7 ETM+ image which also acquired at year 2013.

According to the zoning ordinance of Marion County, Indiana, there are72 types of zoning districts, which were reclassified by permitted land uses and human activities. This study reclassified original zoning districts into eight categories by grouping them by the similarity in land uses and resulting human activities, including Central Business District (CBD), commercial, industrial urban, industrial suburban, urban dwelling, suburban dwelling, agricultural dwelling, and mixed-use district. Zoning districts such as airport, speedways, cemetery, sanitary landfill, power substation were excluded from the analysis, even though the imperviousness and surface temperatures of these zones were relatively high, these land uses have little significance in human settlement. Description of eight planning zones considered in this study is summarized in Table 2.

The Normalized Difference Vegetation Index (NDVI) is a function of the visible and near-infrared reflectance from plant canopy, the reflectance of the same spectra from the atmospheric reflectance [18]. As this index provides an estimate of the abundance of actively photosynthesizing vegetation, thus it can be used to infer general vegetation condition [23, 26, 30]. NDVI is computed as follows: where TM3 and TM4 refer to the reflectivity of Landsat TM band 3 and band 4, with the wavelength of 0.63–0.69 and 0.76–0.90 µm respectively [19, 29]. The value varies from − 1 to + 1, and the greater the value, the better the vegetation condition. NDVI was calculated using ENVI 5.3 software. Area of parcels, building footprints, average building heights, and maximum building heights were calculated using ArcGIS 10.4.1 and ENVI 5.3 software. Percent tree canopy and percent developed imperviousness were derived from National Land Cover Database (NLCD) 2011 by US Geological Survey (USGS).

Subject of socioeconomic vulnerability to extreme heat-related weather event and heat-related death was intensively documented by previous vulnerability studies [12‐14, 16, 31‐34], and many of the previous studies commonly addressed characteristics such as population density, proportion of youth and elder population, female population, non-white population, total housing units, total rental units, total mobile homes, education level, and indices that can explain household economic statuses such as median household income and median house value. Based on the findings of previous empirical studies, this study used aforementioned socioeconomic variables available from American Community Survey 2013. These selected variables were collected at the level of Census block group, which represents the smallest geographic unit for which the desired information was available (Table 1).

To generate the dataset, all of the metrics described in Table 1 were integrated into the land parcel. As a parcel is commonly acknowledged as an important spatial unit of contemporary urban planning and design [35], this study used a parcel as a unit of data integration and as a unit of analysis. Since the average parcel size of Marion County, Indiana was 1526.4 m², each parcel will have approximately 1.7 number of 30-m by 30-m pixels in average. In the case of large parcels, which may contain many pixels, it is possible to have spatially varying surface characteristics such as surface temperature, percent developed imperviousness and percent tree canopy within a parcel. It will be problematic because if a single value, such as percent impervious cover of a pixel, is allocated to a parcel, then spatially varying characteristics that computed in micro-scale will not be counted. In that chase, lots of valuable information measured in a grid level will be lost. To prevent this information loss, this study allowed duplication. Simply put, a parcel may be recognized as multiple observations depends on its size and also depends on the number of pixels observed within a parcel. Spatially varying urban surface characteristics were input into a parcel, resulting in multiple observation which shares macro-level socioeconomic characteristics but has different micro level physical characteristics. In Fig. 3, the example of a Census block group that contains residential, industrial, and commercial zones, multiple land parcels, building footprints, and raster grids is illustrated. In Marion County, there are total 335,489 parcels, and when allowing duplication, then the total number of observations were increased to 942,193, reflecting the number of pixels in the study area. This study divided total 942,193 observations into eight planning zones which share similar urban land surface characteristics. By dividing the original dataset into eight subsets, this study could increase its processing speed and efficiency.

With the advancement of analytical techniques, big data empowers planners and planning decision-makers by helping them to better understand current situations and to predict future more precisely and accurately [36]. Machine learning is one of the proven analytic tools to harnessing the power of big data. It is one of the analytical approaches that uses computer algorithms to repeatedly learn from the data. As the analytical tool continues to learn from the data, the prediction accuracy of the model improves over time [36]. Because the main purpose of this study is in identifying highly determinant characteristics of urban area in the formation of UHIs, this study used the RF method as the main analysis method.

The RF is a classification and regression technique introduced by Breiman [37], and this method is extremely useful in the study highly determinant variable selection because it provides variable importance measure as a part of the analysis results [37, 38]. The RF has been demonstrated to have improved accuracy in comparison to other machine learning methods and also, it is unexcelled in accuracy among current algorithm and it also runs efficiently on large dataset [37]. Moreover, it gives estimates of what variables are important in the classification and it is the main reason that this study used the RF methods over other algorithms to identify the relative importance of urban physical and socioeconomic characteristics to the formation of spatially varying land surface temperature across Marion County. The classification and regression trees are series of binary rule-based decisions that dictate how an input variable is related to its dependent variable. A forest is a collection of trees, and a RF consists of a collection or ensemble of simple tree predictors, each capable of producing a response when presented with a set of predictor values. A RF is random in two ways: (1) each tree is based on a random subset of the observations, and (2) each split within each tree is created based on a random subset of candidate variables [39]. When RF constructs a tree, an average of 36.8% (approximately 1/3rd of cases) of the observations are not used for any individual tree, and these hold-out cases are referred to as out-of-bag (OOB) [37, 39, 40]. In an OOB method, the errors from data excluded for each regression tree generation are used to inform RF of the relative strength and correlation of that tree [37]. This OOB data is used to get a running unbiased estimate of the classification error as trees are added to the forest. It is also used to get estimates of variable importance. The variable importance measure of a variable can be explained as the contribution variables make to the construction of the tree. For a fixed number of trees, a variable with a larger importance score relative to other variables indicates that the variable is important for classification. Therefore, rather than estimate a specific relationship between the independent variables and the response as in data modeling, the variable importance measures are robust statistics pertaining to a variable’s importance in the RF’s emulation of the natural mechanism behind the data [41].

This study used RF add-on package with R statistical software. The R software offers several options for the RF model and its outputs. Among these, the variable importance plot (VIP) and percent increase in mean squared error (MSE) provide relative importance of the independent variables in the prediction of the dependent variable. To calculate percent increase in MSE and produce a VIP plot, the RF algorithm estimate the importance of each predictor by computing how much the error increases for a given tree when OOB data for each predictor are randomly permuted while all other predictors are left unchanged [38, 40]. Only two user-specified parameters are required to run RF: the number of trees in the forest, ntree, and the number of variables randomly sampled at each split, mtry [42]. Following Liaw and Wiener’s [38] suggestion, this study operated RF with a default value of ntree 500 and that of mtry that is the square root of the number of variables in the dataset.

There is no universally accepted variable selection strategy for RF and suggested selection criteria from previous studies are limited [43]. Díaz-Uriarte and Alvarez de Andrés [44] suggested eliminating 20% of the variables having the smallest variable importance and building a new forest with the random variables. They stated the proportion of variables to eliminate is an arbitrary parameter of their method and does not depend on the data. This study followed the strategy proposed by Díaz-Uriarte and Alvarez de Andrés [44], eliminating 20% of variables having the smallest variable importance.

There are two objectives for variable selection. The first is to identify all the important variables, even with some redundancy, highly related to the dependent variable for explanatory and interpretation purpose, and the second is to find a sufficient parsimonious set of important variables for good prediction of the dependent variable [39, 42]. To achieve these two objectives, this study selected variable with RF method then validate its variable selection results with that of the PCA method.

The PCA is a non-parametric mathematical procedure that transforms a number of correlated variables into a smaller number of uncorrelated variables called principal component. The purpose of the PCA is to reduce a large set of variables to a smaller set that still contains most of the information in the large set without redundancy [45]. Thus this study conducted the PCA to identify principal components of the dataset and to use the PCA selection result as a verification tool for the variable selection results by rule-based machine learning technique whether the principal components are successfully identified as important variables. This study used SAS 9.4 software to identify principal components and the correlation between the principal components and the original variables.