How to bring UHI to the urban planning table? A data-driven modeling approach

identified for each feature. The optimal model based on accuracy and interpretability is a decision tree (DT), with an accuracy of 93 %. With the results of the case study, it is demonstrated that (1) the proposed methodology leads to an easy-to-use tool that can be implemented by urban planners to investigate the impact of their design choices at the street-level, and (2) the results obtained are consistent with the current body of knowledge, which in turn alleviates the drawbacks of traditional methods.


Introduction
The Urban Heat Island (UHI) phenomenon is understood as the difference in the temperature between urban areas and the surrounding rural areas caused by the replacement of the natural land surface with the urban fabric (Howard, 1833;Oke, 1982).Howard describes four causes for the differences in temperature in the canopy layer as (1) anthropogenic sources of heat, (2) the geometry of urban surfaces trapping the radiation and blocking its reflection back to the sky, (3) the effect of urban "roughness" impeding the passage of the "light winds of summer," and (4) the availability of moisture to evaporate.UHI has a significant negative impact on health, lifestyle, energy consumption, and greenhouse gas emissions (Akbari et al., 2015).The United Nations estimates that by 2050, 68 % of the world's population will live in urban areas (United Nations, 2019).This rapid urbanization progressively aggravates the effects of UHI, which makes it a priority to consider strategies for UHI mitigation in climate-proof urban development (Parsaee, Joybari, Mirzaei, & Haghighat, 2019).
Mainly, two lines of research have been developed to quantify the UHI, (1) the UHI of the canopy layer is determined by measuring air temperatures, typically at about 2 m above the ground (Stewart, 2011), and (2) the Surface UHI (SUHI) is derived from remote sensing data (Voogt & Oke, 2003).While the first approach, i.e., UHI in the canopy layer, requires the direct measurement of air temperature above the ground refer, the second approach can use remotely sensed data based on the thermal emissivity of land surfaces to measure land surface temperatures (LST).Remote sensing data and above-ground air temperatures are not identical but are correlated (Mostovoy, King, Reddy, Kakani, & Filippova, 2006;Prihodko & Goward, 1997, Zhang, Imhoff, Wolfe, & Bounoua, 2010).
Since the accuracy of the air measurements is influenced, to a large extent, by local conditions (e.g., weather station, observational networks, in situ sensors, etc.), it is difficult to rely solely on these measurements to gain information about UHI at an urban level (Peng et al., 2012).SUHI, however, due to the ability of remote sensing sensors to measure land temperature, provides sufficient data resolution to characterize inner urban centers and rural surroundings.Thus, the terms have become interchangeable in describing UHI patterns.
Surface temperature-based UHI has been successfully used to describe UHI around the world (Shastri & Ghosh, 2019), and its underlying temperature differential within the urban environment (Chakraborty, Hsu, Manya, & Sheriff, 2020;Huang & Wang, 2019;Imhoff, Zhang, Wolfe, & Bounoua, 2010;Li et al., 2018;Liu et al., 2021;Martin, Baudouin, & Gachon, 2015;Streutker, 2002;Zhang et al., 2010;Zhou, Zhao, Liu, Zhang, & Zhu, 2014;Zhou, Rybski, & Kropp, 2017).In this regard, Zhou et al. (2014), by using LST data, investigated 32 Chinese cities, finding that the annual average surface UHI intensity during the day was higher than at night.The authors shed valuable insight into the assessment of the UHI on a regional scale in China.Moreover, Chakraborty et al. (2020), developed an algorithm to map UHI intensity worldwide based on LST.The researchers found, among others, that vegetation can be used to decrease UHI intensities in cities vulnerable to heat-stress.This research follows this trend as well and uses relative LST as an indication of UHI.From this point onwards, relative LST will be referred as the temperature differential within the urban environment.
Of these three categories, the socio-economic and urban morphology factors are of particular interest owing to the fact that urban planning and design decisions made around these factors have a more immediate impact on the mitigation of UHI (Parsaee et al., 2019;Rizwan, Dennis, Chunho et al., 2008).
For politicians and urban planners to be able to consider the impact of their decisions on UHI, the use of simulation models are essentials.
Researchers have deployed various physics-based modeling approaches to simulate UHI (Arghavani, Malakooti, & Ali Akbari Bidokhti, 2020;Berardi & Wang, 2016;Grimmond et al., 2010Grimmond et al., , 2011;;Ozkeresteci, Crewe, Brazel, & Bruse, 2003;Petri, Wilson, & Koeser, 2019;Salata, Golasi, de Lieto Vollaro, & de Lieto Vollaro, 2016;Toparlar, Blocken, Maiheu, & van Heijst, 2017;Tsoka, Tsikaloudaki, & Theodosiou, 2018).One of the most prominent physics-based approach to UHI assessment is Urban Canopy Models (UCMs) (Grimmond et al., 2010(Grimmond et al., , 2011;;Kondo et al., 2005;Oke 1988a, b).The Weather Research and Forecasting (WRF) implements UCMs to explore the impacts of urbanization on the regional climate.This approach can capture interactions between the land surface and atmospheric conditions and thus predict the impacts of urban developments on regional climate change.Jandaghian and Berardi (2020) evaluated the efficiency of three different UCMs to characterize the UHI in Toronto, Canada.The researchers concluded that the more complex UCMs (multi-layer models), do not predict air temperature accurately, mainly because these models do not simulate the variability of urban morphology.In essence, there are two main drawbacks to the UCMs: (1) generating accurate and detailed simulations requires databases with three-dimensional geometry of buildings, vegetation, and structures of the built environment.As a result, calculations of temperatures at each node become very expensive in terms of time and computer power.In most cases, the building geometries are often replaced by homogeneous columns of similar buildings; decreasing the accuracy of the calculations and limiting the number of parameters that can be studied; (2) socio-economic factors, such as population density are not considered.This is because UCMs are focused on the thermodynamic process of the urban environment.This results in

Table 1
Factors affecting UHI summarized in three categories.The environmental characteristics are given by the climate conditions, meteorological features, and geographical characteristics.The socio-economic factors are governed by the intricate nature of the urban function, whereas the urban morphology factor are given by the geometric structure and arrangement of the urban elements.isolated simulations that more often than not, do not represent the intrinsic relationship between all urban parameters at a street-level.According to Baklanov, Mahura, Nielsen, and Petersen (2005) in the current state of the modeling, no one approach fits all needs.
Efforts have also been made from a software perspective to support different degrees of UHI analysis (Esri, 2021;Bruse & Fleer, 1998;Nakano, Bueno, Norford, & Reinhart, 2015;Robinson et al., 2007).These implementations provide urban designers and policymakers with a comprehensive tool to formulate their UHI mitigation designs and policies.However, the use of these tools require a high degree of affinity with both the software interface and the underlying thermodynamic process involved, which makes is less practical for urban planners.
Likewise, several efforts have been made to provide decision-makers with tools to design UHI mitigation strategies effectively.For example, Hoverter (2012), provided an analytical tool for policymakers to consider a combination of four predefined mitigation strategies.More recently, Qui et al. (Qi et al., 2020) presented a framework for identifying an optimal combination of UHI mitigation strategies by integrating performance evaluation models, sensitivity analysis and genetic algorithm (GA) optimization into the selection process.Although these tools are very useful for developing mitigation strategies, they do not provide an immediate assessment about the impact of urban design decisions on UHI.The use of mitigation tools would normally require proactive commitment of planners to apply additional strategies for the reduction of UHI effect (e.g., green roofs).It is conceivable that before reaching to a point where urban planners would want to implement dedicated strategies for the reduction of UHI, they would benefit from a tool that can help them assess the impact of their usual urban design decisions on UHI and determine if they can mitigate it by adjusting their decisions.
Although the existing simulation approaches and methods have shed light on the UHI phenomenon and its possible mitigation strategies, they have four major inherent limitations: (1) These approaches are mainly physics-based and are therefore complex and not user-friendly.Because of this, in the majority of the cases, these models cannot be applied by a wide range of decision-makers who have a little technical affinity with complex simulation models; (2) These approaches often take an isolated view of the problem from a specific perspective focusing on a limited number of parameters.In doing so, they do not take a holistic view of the parameters involved in the study of the UHI; (3) The vast majority of the models focus on building blocks (i.e., micro), or city scales (i.e., macro) (Mirzaei & Haghighat, 2010;Yoo, 2018).This would leave a gap for meso-level modeling at the street level.Street-level analysis of UHI is of particular importance for urban planners and designers, especially in the context of existing cities, because a good portion of the daily decisions made at municipalities are at the street-level.The gap in the simulation resolution results, therefore, in the dismissal of UHI considerations for a wide range of urban design decisions; and (4) The data required by these models are often not collected in a structured and systematic manner (Stewart, 2011).With the increased availability of CityGML models, geographically referenced cadastral datasets, LST datasets, and airborne LiDAR datasets, many of the input features needed for these models are publicly available and can be used for UHI modeling.(e.g., building geometry, canopy coverage, population density).However, the full potential of using these datasets has not been yet realized.
On the other hand, recent advances in data mining technologies and machine learning (ML) modeling approaches have made it possible to extract valuable information from a variety of widely available data sources.For instance, Jiang et al. (2004), implemented an unsupervised ML approach to predict the air pollution levels in China.Data-driven methods have also been successfully used in the area of urban building energy, for example, Nutkiewicz, Yang, and Jain (2018) developed a methodology to characterize and model the energy performance of buildings at multiple spatial and temporal scales by integrating data-driven modeling techniques and building energy simulation approaches.The results show a significant opportunity to improve the accuracy of energy models by integrating ML approaches into the existing simulation workflow.Despite the potentials of the data-driven methods, to the best of authors' knowledge, only a few recent studies have used them to model UHIs (Gobakis et al., 2011;Kwak, Park, & Deal, 2020;Sun, Gao, Li, Wang, & Liu, 2019;Vulova, Meier, Fenner, Nouri, & Kleinschmit, 2020;Yoo, 2018).Vulova et al. (2020) for example, developed a supervised ML model based on the open-source datasets from Berlin to investigate the impact of temperatures in the night hours on the energy consumption at the urban level.Similarly, Kwak et al. (2020) presented an evaluation of sustainable planning in terms of UHI variations through unsupervised learning approaches and statistical methods.
Based on the above review, it can be concluded that there is great potential for the applicability of ML approaches to understand the interaction between different urban parameters.To date, applications to predict temperature variations have been successfully carried out.However, a very limited number of studies have tried to provide decision-makers and urban planners with outcomes that can be easily applied.The trend in that research line places its main focus on prediction and accuracy, rather than on explicability and interpretability.Furthermore, socio-economic features, such as population density or traffic flow are not always included, even though ML approaches can handle a large number of input features.Finally, it is not often that the existing approaches rely on publicly available datasets, which might hinder their dissemination and applicability.
Given the above limitations, the primary goal of the research work presented in this paper is to develop a user-friendly decision-support tool for the assessment of the impact of urban planning decisions on UHI at the street level using a data-driven approach that adopts a holistic view of the socio-economic and urban morphology factors.In doing so, the purpose of this research, and the tool thereof, is not to merely predict the temperature differential at a given geographical location, but rather to provide an explanation about the root causes of the temperature differential in terms of socio-economic and urban morphological factors.This tool will offer an easy-to-use classification of the streets in terms of UHI, based on the factors that are within the design space of urban planners.As shown in Fig. 1, the envisioned tool will only be used by the end-users (e.g., urban planners, decision-makers, etc.) once it is fully developed.This is to say that a system developer will collect the relevant data and develop the decision support tool that can then be used by the end-user.As such, the end-users are not concerned with the development phase of the tool.
By pursuing this objective, therefore, this research aims to make the following contributions to the body of knowledge: (1) to propose a systematic approach for the collection of data that can be used in UHI simulation; (2) to develop an easy-to-use tool that can be used by urban planners to investigate the impact of their design choices on UHI at the street-level; and (3) to bridge the gap between micro and macro-level simulation of UHI.The remainder of this paper is structured as follows: first, the proposed method is explained in detail.Followed by the presentation of a case study for the city of Montreal, Canada.Finally, the discussion and conclusions are presented.

Proposed method
To address the research gap presented in Section 1, a methodology is adopted to develop the decision support tool for the modeling of UHI at the street-level.As shown in Fig. 2, this methodology consists of four phases: (1) Data collection, (2) Data preparation, (3) Model development, and (4) Development of the UHI Assessment Matrix.In a nutshell, the data collection phase aims to use the publicly available datasets to collect relevant socio-economic and urban morphology factors.The data preparation phase involves the extraction of the features from the raw data collected in the previous phase.In phase 3, i.e., model development, the processed data are studied to determine the features that can best explain UHI at the meso-level by using an optimization-based approach.Lastly, for the development of the UHI Assessment Matrix, a tree-regression approach is applied to elaborate a set of street categories that will categorize the type of street by their temperature category.

Data collection
As previously stated, factors affecting UHI at the urban level can be grouped into three main categories, i.e., environmental factors, socioeconomic context, and urban morphology.Environmental factors have been described in the literature as uncontrollable factors, while the socio-economic context, and urban morphology factors, are those that have the greatest potential to be influenced by policies and urban planning decisions (Parsaee et al., 2019;Rizwan, Dennis, Chunho et al., 2008), and as such, they are the focus of this research study.Table 2 provides an overview of the dominant features affecting UHI.In this table, primary features refer to those features that need to be directly measured, while secondary features can be derived from primary   features.For example, impervious surfaces, which are often referred to in the literature as artificial water-resistant structures such as asphalt or concrete, are considered to be the main contributor to the UHI effect (Imhoff, Zhang et al. 2010).This feature, however, is not independent, as its value depends on the built-up ratio, vegetation, and water densities.Its calculation as shown in Eq. 1 is rather straightforward.The impervious surface is equal to one minus the total sum of the built-up ratio, vegetation, and water densities.
Where: IS: Represents the impervious surface density BD : It is building density VD : It is vegetation density WD : It is water density Transpiration from buildings and infrastructure, likewise, is an immediate result of both the use and emission of heat from buildings and urban mobility, in combination with the presence or absence of vegetation (Magli, Lodi, Lombroso, Muscio, & Teggi, 2015).Similarly, the Sky View Factor (SVF) is determined by a fraction of the visible sky at a point, which can be understood as a composition of the building height (H) to street width (W) and the presence of other infrastructure or vegetation (Oke et al., 2017).It has been acknowledged in the literature that data-driven techniques best perform by limiting the feature domain to primary features that are not redundant within themselves (Guyon & Elisseeff, 2003).Therefore, to build a compact model with better generalization capabilities, only the primary features will be considered in this study.
There are three major groups of data from which the primary features can be extracted, namely Airborne LiDAR datasets, LST datasets, Population density (defined by the measurement of the number of citizens per unit area) can be perceived as historical data that are usually captured every decade by means of the population census.Statistical projections are usually made each year to give an estimate of the population growth and are usually available in the form of counts per administrative unit.Advances in geospatial modeling techniques make it possible to map the population distributions at a resolution of 3 arc seconds (Sorichetta et al., 2015).
Transportation refers to the traffic flow in urban areas as well as the transportation infrastructure.Transportation infrastructure datasets are stored as a GIS vector dataset and pertain to the transportation infrastructure of the city, e.g., bridges, tunnels, highways, roads.The traffic flow datasets can be retrieved from the city and municipal authorities.This type of data is stored as historical data, traffic count, or real-time traffic data.Surface characteristics and building materials are not usually surveyed by the local governments.They are, however, constrained by design regulations, and therefore limited to a number of options, which makes it possible to conduct a visual inspection of the area of interest.
Regarding temperature data, LST are widely used as an indicator for UHI (Shastri & Ghosh, 2019).These measurements are usually provided in GIS format, from which a temperature index can be extracted for the street segment by calculating the average temperature in each overlapping area.For this study, this temperature differential extracted from the LST will be considered as the label data in the model development.

Data preparation
Data preparation consists of two main input sources as shown in Fig. 3: namely, (1) GIS datasets preparation, and (2) LiDAR point cloud.
To process the GIS datasets, it is necessary to check the correct coordinates of the area of interest and the accuracy of the geometries.This is done by first projecting the GIS layers into the correct coordinate system.This step is followed by extracting only the geometries of interest.Finally, all geometries are validated and repaired based on OpenGIS standards (Herring, 2014).In the case of polygons, for example, some polygons could be self-intersecting, in which case the geometry containing errors needs to be eliminated.This is a very simple step that can be done employing any geo-analytics tool.
To process the LiDAR point cloud datasets, it is necessary to assemble them into a LAS dataset that covers the area of interest.From here, digital terrain models (DTM) and digital surface models (DSM) can be extracted.This is done by filtering the LAS datasets by point classification.Only the ground points should be selected to create the DTM.As for the DSM, only the points representing the vegetation and buildings should be selected.To generate the height of the features, a normalized DSM (nDSM) is processed.This is done by subtracting the DTM from the DSM.Subsequently, the nDSM shows the height of the features above the ground.The heights of the buildings are then calculated from the nDSM within each building footprint polygon.Fig. 4 shows the overall idea of the process.
Before extracting the primary urban features, all features must be scaled down to the street level.To this end, the jurisdiction and the area of interest of each street segment must be defined.The area of interest is the area surrounding the street segment from which the urban features can be assembled.In other words, this is the minimum area from the street profile in which all the features surrounding the street segment can be group into one unit.As shown in Fig. 5, the jurisdiction of each street segment is first determined by finding the intersection of street centerlines with each other.Then, the buffer zone b can be used from the centerline of the street segment based on the area of the average setback defined by each city plus the average width of both the street and the sidewalk.

Socio-economic factors
This section describes the extraction of the socio-economic factors, which include land-use, population density, transportation, and surface and building materials.

Land-use.
Land-use refers to the purpose that the land serves, such as commercial, residential, industrial, or open space, among others.It is important to note that land-use attributes are usually nominal values that need to be encoded to numbers before they can be used.To calculate the land-use distribution per street-segment, each segment must be assigned a land-use category.This is done by considering the most frequent category in the histogram of the land distribution of each street segment.Fig. 6 illustrates the general idea, where for street A the dominant land-use is residential.

Population density.
Population density can be calculated by projecting the administrative unit-based census into geospatial gridded cell datasets and computing the average population per street-segment, as shown in Fig. 7.

Transportation.
Usually, traffic data are collected at major intersections or highways.However, to have a homogenous distribution of traffic data, it is important to translate these point data into a rasterized density value.As shown in Fig. 8, a gridded traffic count can be Fig. 4. From the LiDAR point cloud datasets, the DTM and DSM can be extracted.Next the height of the features can be calculated by taking the difference between the two models (DSM -DTM) within each building footprint polygon.generated from this traffic count point data.To this end, the value of each gridded cell is considered as the total traffic count of all the intersections in that cell.It is worth noting that the resolution of the traffic data needs to be upscaled to increase the homogeneity of the data for the model development.

Surface and building materials.
The characteristics of the surface and building materials are intrinsic to the built environment.The interaction between UHI and façade materials has been of great interest to researchers (Manni, Lobaccaro, Goia, Nicolini, & Rossi, 2019) since the material of buildings' façade is known to play an important role in the formation of UHI.Therefore, façade materials can impact the building's energy consumption, and consequently CO 2 emission at the urban level (Berardi, GhaffarianHoseini, & GhaffarianHoseini, 2014).For the purpose of this study, only surface materials of the façades are considered.
It is important to mention that surface and building materials are categorical features that need to be encoded to numbers before they can be used.The calculation of this factor is done according to the procedure adopted in the case of land-use (Fig. 6), where the most dominant material is obtained from the analysis of the histogram of each street segment.

Urban morphology factors
This section details the extraction of urban morphology factors.They are the building geometry, building height to street width ratio, water bodies/cool sinks, built-up ratio, vegetation/green spaces, and impervious surfaces.

Building geometry.
Building geometry refers to the physical characteristics of buildings, i.e., height, shape, and area.Since this study only focuses on the characteristics by street-segment, the shape of each building is not taken into account.However, the building geometry will be represented in terms of the dispersion of building heights and the average height per street-segment.
The heights of building are extracted directly from the LiDAR point  cloud, where the heights are calculated from the footprint polygons of each building and the nDSM, as shown in Fig. 4. The dispersion of the building heights per street-segment is therefore computed by calculating the standard deviation of building heights per street-segment.

Average building height to street width ratio (H/W ratio). H/W
ratio or the canyon aspect ratio is the ratio between the height of the building to the street width.To calculate the H/W ratio per street segment, the average height of the buildings is divided by the width of that specific street segment, as shown in Eq. 2 and illustrated in Fig. 9.
H/W ratio = Average building heigh in the street Width of the street segment (2) 2.2.2.3.Densities per street-segment.The densities per street segment refer to the calculation of the proportion of buildings, water bodies, also called cool sinks, and vegetation density (green spaces).The areas of overlap of these characteristics with the area of the street-segment are taken to calculate the corresponding densities.Eqs.3-5 show the computation of the built-up ratio, vegetation densities, and water bodies per street-segment, respectively.These densities need to be calculated one by one and then added as independent features to the overall list of features.Fig. 9.The street canyon aspect ratio per street-segment is calculated taking the average height of the buildings divided by the width of that specific street-segment.
M. Pena Acosta et al.

Model development
To reduce complexity and enhance the design of a simple decision support tool for governments and urban planners, it is useful to evaluate the importance of each of the features mentioned in Section 2.1 for the prediction of the intended output (i.e., temperature differential).This information can be used to determine which features need to be included in the model (i.e., feature selection).The goal of the model development is twofold: (1) to find the optimal number of features that are highly related to temperature differential, and (2) to develop the simplest model that maintains the desired accuracy.Given that the feature selection procedure occurs at the same time as the model development process, it can be classified as an embedded methodology.
Fig. 10 presents an overview of the model development.Initially, the datasets are split into training and testing datasets to train the regressors.In this research, given that the temperature data are used as the indexed output of the model, the supervised methods of Decision Trees (DT) and Random Forest (RF) are selected as the main regressors.This is Fig. 10.In the model development the datasets are firstly split into training and testing datasets to train the regressors.Then, the hyperparameters of both regressor are set by default.This is followed by an optimization process in which the minimum number of levels of the tree (max-depth) is selected randomly, while minimizing the number of features that can successfully boost the learning process.Once the performance threshold is reached, the most important features with the max-depth from the best performing models are selected.
because the interpretability of the model is a fundamental trait for this research, as the model aims to be interpreted and communicated to decision-makers explicitly.Therefore, to derive explicit interpretation from the learning process, the selection of the learning approach is reduced to tree-based regression algorithms.
RF and DT are among the most frequently used tree-based algorithms in supervised learning, as they offer training efficiency, high prediction capabilities with minimal hyperparameters tunning, and do not require normalization of features.However, the prediction capacity of the RF is generally higher than that of DT, but in turn, it does not offer the same level of interpretability as the DT does.Researchers are often faced with the trade-off between performance and interpretability (Pintelas, Livieris, & Pintelas, 2020).Therefore, both regressors are proposed in this research to compare performance and interpretability.
A DT intrinsically performs feature selection by choosing appropriate split points.Based on this knowledge, the importance of each feature can be evaluated to only select the most important features.However, in a RF model, as the forest grows, the interpretability of the model decreases.For this particular reason, both DT and RF can be optimized by tuning their hyperparameters, and thus increasing the performance of the models while acquiring better interpretability.In the case of the DT, the optimization helps improve accuracy while maintaining interpretability, which will be achieved by minimizing the number of layers in the tree as well as the number of input features required.
In this research, the interpretability of the model, in particular the readability of the tree, is crucial.Therefore, in the training of the regressor, the effort must be made to find the minimum number of levels of the tree (max-depth) while minimizing the number of features that can successfully boost the learning process.
To this end, a GA-based optimization approach is implemented.An initial set of candidate solutions (the initial population) is therefore randomly selected, based on which the DT and RF regressors are trained.The optimization process aims to minimize the prediction error, which is measured in terms of the Mean Absolute Error (MAE) of the predicted temperature value.Then, mutation and crossover operators are applied to evolve the best individuals and to generate a new population.Given the mixed nature of input variables in this problem, i.e., the binary values for features to include and an integer value for max-depth, a customized structure is developed for the chromosome to prevent the mutations and crossover between binary and integer values.This process is iteratively repeated until the minimum MAE is reached.At the end of the iteration process, the best individuals represent the list of chosen features and the specific depth of the tree with the minimum MAE.Fig. 11 presents some of the main ingredients of the GA-based optimization process.

Development of UHI assessment matrix
While the models developed in the previous phase can potentially be used to predict LST variation within a given urban area based on the extracted features, by their own, they would not allow a direct interpretation of the rationale behind the predictions.This is because, as discussed in Section 1, ML approaches are known to be black boxes, with limited to no explanation of the predictions.A DT, however, under a manageable number of features and layers in the tree, offers the possibility to examine the interplay between influencing variables and output predictions.
Hence, to facilitate an explicit interpretation of the models, an assessment matrix of the UHI is developed based on an in-depth analysis of the decisions made by the DT at each node to reach the final predictions (Fig. 12).
The UHI assessment matrix aims to create a categorization of the street types according to their temperature label based on the DT model.Fig. 13 shows an example of how the prediction at the leaf node is reached.Since each decision has a feature associated with it, it is possible to predict based on a set of sequenced thresholds of features.

Fig. 12.
For the development of the UHI evaluation matrix, the street temperatures are first classified into a 5-scale range.From here the selected regressor is trained with the optimal configuration determined in the previous phase.Once the model is trained, the set of rules leading to each temperature category can be analyzed and extracted.In the final step, the matrix can be developed based on a set of sequenced feature thresholds.
In this way, the predictions (temperature differential) will become the categories to which each feature will be assigned.Starting from the root node, the decision path can be examined node by node.Within each node, a threshold of how the splitting was made can be extracted.Once all the leave nodes have been reached, the threshold of each of the above nodes can be grouped in a category, as illustrated in Fig. 13.For instance: Temperature 1 = feature 1 < feature 2 AND feature 1 < feature 3 AND feature 1 < feature 4 Temperature 4 = feature 1 < feature 2 AND feature 2 < feature 3 AND feature 2 < feature 5 This process can be manually repeated for all the leaves nodes in the model.

Case study
To evaluate the effectiveness and feasibility of the proposed methodology, a case study based on the city of Montreal was selected.Montreal is located in the south-east of Canada and is its second most populated city.Covering an area of approximately 499 km 2 and a population of 1.9 million inhabitants, it has a population density of 3.9 inhabitants/km 2 (Statistics Canada, 2016).The city has a continental climate with cold winters averaging -10 • C and hot summers between 22 and 32 • C. Montreal is considered a metropolitan city, with a wide-ranging urban geometry (Fig. 14), from downtown with high-rise buildings (the tallest of about 220 m), to suburban areas featuring flat buildings with an average height of 20 m.Such a variety of urban scenarios provide the model with a good basis for a data-driven analysis of the configuration of different urban features in terms of the UHI.
As discussed in Section 2.1, publicly available datasets were retrieved from the Montreal Open Data Portal (Open Data Portal, M., 2021).These datasets contain airborne LiDAR datasets, GIS datasets consisting of the vector layers of the city's land-use distribution, road network, traffic counts per major intersection, building footprints, and the canopy coverage.Additionally, the population density was retrieved from WorldPoP (WorldPop Data (2019a), Bondarenko, Kerr et al., 2020).Regarding the temperature data, due to the accessibility of the datasets, relative LST was obtained in raster format from Partenariat Données Québec (2012).

Data inputs
The LiDAR dataset consist of a large number of 3D points recorded (up to 400,000 per second), providing information about ground elevation, and features above ground, and the presence of dense vegetation.The point cloud data were collected employing Airborne LiDAR (XEOS) from 2005 to 2011.The points were classified from 1 to 8 as presented in Table 3.
Land distribution compromised the geometric representation of the land allocation.Each category was assigned a numerical value that represents one of these land-use types.Table 4 summarizes the types of land-use and their corresponding numerical labels.
Regarding the road network dataset, it contains a GIS layer representing the road network.Each street was originally segmented by its intersection with other streets.Fig. 13.Overview of the decision-tree categories extraction At each node, a threshold of how the splitting decision was made is extracted.In this way the final predictions (the temperature label) can be determine based on set of sequenced thresholds of features.For example, Temperature 1 = feature 1 < feature 2 AND feature 1 < feature 3 AND feature 1 < feature 4.This will therefore, become category 1: UHI potential low.Building footprints, canopy coverage, and water presence, likewise, were also 2D representations of the area occupied by these infrastructures, vegetation, and water across the island.However, building footprints were only available for the boroughs of Ahuntsic-Cartierville, Outremont, and Ville-Marie.
In this regard, the building heights were extracted as described in Section 2.2.2.1.To ensure the accuracy of the data obtained from the nDSM, the resulting heights were manually checked against the measurements from Google Earth.25 buildings with different heights from different parts of the city were considered.The average estimation error was 6% with a standard deviation of 0.03.This indicates that the estimation error is marginal, and thus the data was accurate enough to be used in the case study.
As for the temperature input datasets, the relative LST data was obtained in a rasterized GIS format.This temperature data was calculated with images from the SPOT-5 satellite and surface temperature readings from the Landsat-7 satellite.A total of 67 images were collected from 2005 to 2011, always during the summer months.The methodology implemented by CERFO relies on a mathematical model to combine surface temperatures from low-resolution readings (100 × 100 m per pixel) into high-resolution images without thermal data (20 m).The map does not show absolute temperatures, but rather relative  temperatures on the island, representing coolest areas (minimum relative value of 1) and the hottest areas (maximum relative value of 9) as shown in Fig. 15.The methodology implemented to generate these relative measurements can be found in (Boulfroy, Khaldoune, Grenon, Fournier, & Talbot, 2013) As no data on the façade materials were available, a visual inspection was carried out via Google Earth.This was only done for the boroughs where the building footprints were available.As a result, five dominant materials were identified, namely, concrete (precast), glass, masonry (bricks/stones) -grey to reddish, masonry (bricks/stones) -red to brownish, and wood.Buildings were assigned a numerical value that represents one of these types of materials.Table 5 lists the types of materials and their corresponding labels.
The population density was retrieved in a rasterized GIS format at a resolution of 3 arcs.The units are the number of people per pixel.The methodology used and the sources of the input data are described by (Bondarenko, Kerr, Sorichetta, &Tatem, 2020), and(WorldPop Data (2019).Transportation-related information per major intersections was obtained from (Open Data Portal, M., 2021).It consists of the number of cars, trucks, buses, and motorcycles registered in peak hours during the period from 2008 to 2015.As presented in Section 2.2.1.3,these transportation counts were summarized by intersection ID and rasterized to a 1 km 2 resolution.

Data preparation
Street-level features were calculated using ArcGIS Pro 2.5.2, the 3D Analyst, Spatial Analyst, and Geostatistical Analyst extension tools (ArcGIS Pro | 2D and 3D GIS Mapping Software).To build the data instances, first, a buffer of 25 m (based on the average setback defined for the island, plus the average street and pedestrian width) was defined around the centerline of each street.From here the socio-economic and urban morphology features were extracted.Regarding the water density, there was no significant presence of water bodies or cool sinks in the areas of interest, therefore for the case study, this feature was not included.Table 6 presents a summary of the feature extracted by their corresponding data type, mean, minimum, and maximum values.A total of 5.578 data instances were extracted.

Model applicability and validation metrics
An overview of the total dataset used is presented in Table 7.The dataset is structured in such a way that the columns represent different features by category and the rows represent the data instances.The dataset was divided into training (75 %) and testing datasets (25 %).Specifically, 4,183 feature vectors were used to train the models and 1,395 feature vectors were used to validate the performance of the models.
The implementation of the RF and DT algorithms was carried out through the ML python library scikit-learn (Pedregosa et al., 2011).The algorithms were initialized with the default hyperparameters configurations, except the maximum depth.The final list of features and the maximum depth of both algorithms (i.e., RF and DT) were optimized by selecting the least number of features and the minimum depth of the regressors while minimizing the MAE.
For the DT, the initial population was set as 700 and evolved over 100 generations, with a probability of mating two individuals (CXPB) of 0.8, and a probability of mutating an individual (MUTPB) of 0.2.All features were used as input data for the optimization process.The results show that a higher accuracy of the algorithms was reached by only considering building and vegetation density, average building height, average traffic count, predominant façade materials, and land use, with a maximum number of layers (max-depth) in the tree of 7.
Regarding the FR, the initial population was set to 100 and evolved over 50 generations, with a probability of mating two individuals (CXPB) of 0.8, and a probability of mutating an individual (MUTPB) of 0.2.Higher accuracy was achieved with a max-depth of 10 and with the same features as the DT model but also including the building standard deviation.
Fig. 16 presents the dispersion of the predictions in terms of the predicted temperatures of both regressors.The optimized RF shows better accuracy and generalization capabilities, with lower data dispersion and higher precision.The DT, however, shows more variability (higher dispersion).2017)).Numerical label refers to the label that was assigned to each material.Although the optimized RF model shows better predictive performance (93.81 %) that can be explained by its nature of choosing features randomly during the training process, and thus, it does not depend heavily on the specific set of features, the complexity of the model makes its interpretation more difficult.The DT, on the other hand, performs efficiently with a prediction accuracy of 93.03 %.However, given its better interpretability, the DT was selected to be incorporated in the decision-support tool to aid the decision-making process.

UHI assessment matrix
By examining each node and its consequent path from the trained DT model, a set of thresholds were extracted that represent five categories, corresponding to five ranges of UHI, from low to high.Table 9 summarizes the five categories and their respective thresholds.
Category 1 represents low UHI potential -L.This category is defined by a low density of buildings, with a low average height of buildings per street segment.Parks and conservation areas prevail, with a high density of vegetation.Both the population density and the traffic flow are low.
Likewise, category 2 represents medium-low UHI potential -ML.It is defined by the low built-up density, and a slightly lower density of vegetation.Here, it is already possible to see that as the vegetation density begins to decrease, the temperature starts increasing.The average building height, as opposed to what is expected, is lower than that of the previous category.In this category, however, more residential buildings can be found, and therefore, the population density increases as well as the traffic flow.
Category 3 refers to medium UHI potential -M.Similarly, to the previous category, the land-use is predominantly dedicated to residential purposes.The population density is higher, and the traffic flow is greater.An increase in built-up to 32 % and a 20.8 % decrease in vegetation density are also observed.
Category 4 represents medium-high UHI potential -MH.The temperature in this category is already high and the built-up density remains constant.However, the vegetation density continues to decrease.The predominant land use in the category is dedicated to residential, commercial, and industry, which in turn increases the flow of traffic.Fig. 16.Models accuracy in terms of the predicted temperatures of both regressors.The RF shows better accuracy and generalization capabilities, with lower data dispersion and higher precision.The DT shows higher dispersion.However, the population density remains the same as in the previous category.
Finally, Category 5 represents high potential UHI -H.In this category the build-up density covers a range of 32 %-51 %, although the vegetation density remains similar to the previous category.The average building height is the highest and the land-use falls predominantly into the commercial and industrial categories, with concrete and glass as predominant façade materials.

Discussion
With the uncertainties that climate change brings to our societies, it becomes even more essential to provide urban planners and decisionmakers with state-of-the-art and user-friendly methodologies to incorporate UHI considerations in their designs.While many studies use additional data collection procedures, such as sensors, satellite monitoring systems, or physics-based simulation, the methodology proposed in this paper uses publicly available data to develop accurate models that can be easily translated to easy-to-use decision-making assessment matrix.It is shown that urban morphology characteristics and socioeconomic features can be easily extracted from this data.It is also shown that these data can explain the variation in temperature in the city with high accuracy.
A large set of urban features were extracted from which it was possible to gain a deeper understanding of UHI and its relationship with a large number of urban morphological and socio-economic factors.For instance, the results show that vegetation density is the most important feature and there is a clear correlation between the increase of temperature and the decrease of urban vegetation across the studied areas.This result is consistent with previous research (Yoo, 2018).However, features such as traffic count and average building height depend on their interplay with the other urban features to have an impact on the temperature.This emphasizes that while the methodology for extracting UHI knowledge from already existing datasets is the same, the relationship between the feature hinges on the city's socio-economic context, and thus a holistic approach should be applied when studying the UHI phenomenon.
In addition, as discussed by Parsaee, Joybari et al. (2019), the success of UHI mitigation strategies depends largely on the decisions made by urban planners and politicians, and thus, the explainability and interpretability of the models are fundamental.The knowledge gained by modeling all these features and relationships can be effectively transferred to those who can apply the results.In this regard, Yoo (2018) identified the most important features by implementing an RF and principal component analysis (PCA) approach.These two methods, however, are not well-suited for a straight-forward interpretability of the predictions.On the other hand, in the present study an interpretable modeling approach (i.e., optimized DT) was implemented, from which insights were extracted and translated into an UHI assessment matrix.In this way, decision-makers can easily identify urban features by their corresponding UHI category.
Regarding the applicability of ML algorithms to assist urban studies, the results presented in the paper attest to the potentials of data-driven approaches for overcoming the limitations of physics-based models in creating an accurate and consistent understanding of the UHI.As highlighted by Santamouris, Synnefa, and Karlessi (2011), a deep understanding of the UHI plays an important role in overcoming its negative impacts on societies.The authors can envision that in the long term, when sufficient datasets are collected from various urban environments worldwide, a meta model can be developed to explain the UHI phenomenon not only with respect to the socio-economic and morphological feature, but also the geographical location.
The results of the present research show that not only is it feasible to explain the UHI by the implemented ML algorithms, but also that it is concomitantly possible to gain a deeper understanding of the phenomena, and thus, alleviating the complexity of physics-based modeling while extracting knowledge from publicly available datasets.

Conclusion and future work
This research work presented a methodology for the development of a decision support tool for the street-level modeling of UHI.This is an easy-to-use tool that can be used by urban planners to investigate the impact of their design choices at the street-level.By using publicly available data from the city of Montreal, the applicability of the proposed methodology was demonstrated.The UHI assessment matrix represents five levels of UHI potentials, ranging from low to high.This is a straightforward categorization of different urban features by their UHI category, which is rather easy to interpret and inexpensive to build, as it does not require additional means of data collection.
The presented approach takes into consideration socio-economic and urban morphology factors at a street-level, presenting a holistic view of the parameters involved in the study of the UHI.By doing so, it bridges the gap between micro (i.e., building blocks) and macro (i.e., city) scale studies.Additionally, by scaling down urban features to a street-level analysis, the proposed methodology offers the possibility to design mitigation strategies based on factors that are within the design space of urban planners.
Although the presented methodology has proven to be efficient to derive easily interpretable knowledge on UHI from publicly available datasets, there are three main points of improvement: (1) The data used to correlate the façade materials was obtained by visual inspection.In future studies this data collection can be performed by automatic means of classification and image detection.In addition, it should be highlighted that the contribution of the façade system to UHI is governed by the complex heat transfer characteristics of the building envelop.Nonetheless, because the focus of this research was on the applicability and ease-of-use of the developed decision support tool for urban planners, this research tried to simplify this matter by making the assumption that the type of the dominant material in the street can be used as an approximation for the contribution of the façade system to UHI at the street level.Having said that, the model can be enhanced in the future by including more detailed data on the specificity of the complex behavior of façade systems (i.e., material thickness, layers, etc.) mainly at the building level.
(2) DTs are known to be susceptible to variance errors.
Although the model implemented in this research showed low variance, it is important to note that the specific values extracted from the regressor may differ under a different sets of input features.Thus, it is worthwhile to apply the proposed methodology to other urban settings in subsequent studies.Additionally, as the urban area may evolve over time, the UHI assessment matrix may need to be updated as well to reflect these changes in the built environment.However, considering the rate of urban growth and the magnitude of changes that may justify the reevaluation of the assessment matrix, this is only expected to happen once every few years.The methodology proposed in this research offers a clear and easy-to-follow process to do this update easily.
(3) Although SUHI has been successfully used to describe UHI worldwide, and its underlying temperature differential within the urban environment, an air temperature-derived UHI assessment matrix could potentially be different, as canopy temperatures and LST show slightly different intraurban thermal patterns.As such, with the availability of more air temperature measurements across the built environment, a more complete dataset could easily be incorporated into the proposed framework.This may lead to a more air temperature-driven UHI Assessment Matrix that can be slightly more relevant to public health issues.This can be explored in future studies.

Fig. 1 .
Fig. 1.Sequence diagram of developer and end-user interactions with the UHI Assessment Matrix development process.In the envisioned tool, the developer will collect the relevant data and develop the decision support tool that can then be used by the end-user.Therefore, end-users are not involved in the development phase of the tool.

Fig. 2 .
Fig. 2. The proposed method comprises four main stages.In the first stage, raw data are collected from publicly available datasets.Then, relevant features are extracted in the data preparation step.The third stage deals with the model development, in which the features are studied in detail.Finally, in the development of the UHI Assessment Matrix, the classification of street types according to their temperature label is created.

Fig. 3 .
Fig.3.The data preparation process involves the extraction of the relevant socio-economic and urban morphology features from the raw data.This process consists of two main parts: (1) the preparation of the GIS datasets and (2) the preparation of the LiDAR point cloud datasets.Once the datasets are prepared, the street jurisdiction is determined.Then, the relevant socio-economic and urban morphology features are extracted and downscaled to street-level for further analysis.

Fig. 5 .Fig. 6 .
Fig. 5.The determination of street jurisdictions can be defined as the minimum area from the street profile in which all the features surrounding the street segment can be group into one unit.This area is represented by the buffer zone b.

Fig. 7 .
Fig. 7.The calculation of population density per street-segment is done by projecting the administrative unit-based census into geospatial grid cells, so that the number of people per grid cell can be estimated.Then, the average population per street-segment is calculated by taking the average of the number of people in each grid cell.

Fig. 8 .
Fig. 8.To calculate the traffic count per street-segment, first the average traffic count per major intersection is translate into a rasterized density value.Then the value of each gridded cell is considered as the total traffic count of all the intersections in that cell.

Fig. 11 .
Fig. 11.Main ingredients of the GA-based optimization process.The initial population consisting of the initial set of candidate solutions is randomly generated.To perform the mutation and cross-over operations over the generations, a customized chromosome structure is developed to avoid mutations and crossovers between binary (features) and integer (max-depth) values.

Fig. 14 .
Fig. 14.Examples of different urban geometries in Montreal with wide-ranging urban geometry: (a) business and governmental areas in Ville-Marie, (b) suburban area of Outremon characterized by flat buildings with an average height of 20 m; and (c) high-rise buildings (the tallest of about 220 m) in Ville-Marie.

Fig. 15 .
Fig. 15.Relative LST retrieved fromBoulfroy et al. (2013).Where the Island of Montreal was mapped to produce relative temperatures, representing cool areas (minimum relative value of 1) and the hottest areas (maximum relative value of 9).

Table 2
Overview of the dominant features affecting UHI.Primary features refer to the features that need to be directly measured.Secondary features can be derived from primary features.
GIS vector files, LiDAR databaseAverage building height to street width ratio Built-up density Water bodies/cool sinks Vegetation/green spaces M. PenaAcosta et al.

Table 3
LiDAR points classification.

Table 4
Summary of land use types and corresponding numerical labels.

Table 5
Summary of the five most dominant materials found in the boroughs of Ahuntsic-Cartierville, Outremont, and Ville-Marie, with their corresponding albedo and emissitivity values.The values are given as a range or average.

Table 6
Summary of the feature extracted by their corresponding data type, mean, minimum, and maximum values for each category.

Table 7
Overview of dataset used to train the DT and RF regressors.In total, 5.758 instances were extracted from the raw datasets.

Table 8
GA-optimization parameters and results for both DT and RF regressors.The crossover (CXPB) and the mutation (MUTPB) probability values were considered the same for both regressors.However, the initial population for DT comprised 700 individuals which evolved over 100 generations.For the RF, a population of 100 individuals and 50 generations were considered.The RF showed a mean absolute error (MAE) of 0.41 and an accuracy of 93.8 %, while the DT performed with a MAE of 0.47 and an accuracy of 93.03 %.

Table 9
UHI Assessment Matrix presents a set of thresholds that were extracted from the trained DT model categorized by UHI potential.Each category represents five ranges of UHI, from low to high.