Fine-Scale Urban Heat Patterns in New York City Measured by ASTER Satellite – The Role of Complex Spatial Structures

Urban areas have very complex spatial structures. These spatial structures are primarily composed of a complex network of built environments, which evolve rapidly as the cities expand to meet the growing population’s demand and economic development. Therefore, studying the impact of spatial structures on urban heat patterns is extremely important for sustainable urban planning and growth. We investigated the relationship between surface temperature obtained by the Advanced Spaceborne Thermal Emission and Reflection Radiometer (ASTER, at 90 m spatial resolution) on the current EOS-Terra platform and different urban components based on the classification of high-resolution QuickBird imagery. We further investigated the relationships between surface temperature and building footprint and land use information acquired from the New York City (NYC) Department of City Planning. The ASTER image reveals fine-scale urban heat patterns in the NYC metropolitan region. The dark and medium-dark impervious surfaces, along with bright surfaces, generate higher surface temperatures. Even with highly reflective urban materials, the presence of impervious materials leads to an increased surface temperature. At the same time, trees and shadows are effective in reducing urban heat. The data aggregated to the census tract reveals high-temperature clusters in Queens, Brooklyn, and the Bronx region of NYC. These clusters are associated with industrial and manufacturing areas and multi-family walk-up buildings as dominant land use. The census tracts with more trees and higher building height variability generate lower surface temperatures, consistent with shadow cast by high-rise buildings and trees. The results of this study can be valuable for urban heat island modeling on the effects of building heights variability and tree shadows on small-scale surface temperature patterns. It can also help identify the risk areas during extreme heat events to protect public health.

and prevent heat release back to the atmosphere. The UHI can be complicated by the land-sea breeze's cooling effect (Yoshikado, 1990(Yoshikado, , 1992. Complex urban geometry casts shadows between high-rise buildings, reduces sun exposure, and narrows the surface view of the atmosphere, resulting in a lower surface temperature (Nichol and Wong, 2005;Guo et al., 2016).
Anthropogenic factors, such as winter heating or summer cooling systems from commercial and residential buildings, amplify the intensity of the UHIs. Trees in the neighborhood increase the likelihood of cooler surface temperature because of excess heat released by trees through evapotranspiration. The SUHI is typically characterized by airborne or satellite thermal infrared remote sensing at regional scales. The recent developments in spaceborne and airborne remote sensing technology and the availability of thermal sensing data from various satellites makes it possible to study SUHI effectively (Voogt and Oke, 2003;Zhou et al., 2019). Our study focuses on using thermal remote sensing data to examine the spatial structure of urban thermal patterns and their relation to urban surface characteristics.
In big cities, tall buildings often cast shadows on buildings and the ground. Loughner et al.
(2012) observed that higher surface temperature is associated with shorter urban buildings based on the model simulation. They concluded that shorter buildings cast fewer shadows and allow heating of the building walls and roads through direct sunlight (Loughner et al., 2012). On the other hand, taller buildings cast larger shadows and reduces solar radiation absorption onto urban impervious surfaces. The availability of high-resolution imagery from satellite sensors such as IKONOS and QuickBird provides a unique opportunity for detailed urban land cover mapping.
We can obtain accurate urban land use information, including impervious surfaces, vegetation cover, spatial organization of urban buildings, shadows, and surface temperature patterns. The climate is controlled by cold, dry air mass movement from the north and warm, humid air mass movement from the south. In addition to the above two air masses' interactions, the air mass flow from the North Atlantic Ocean produces cool, cloudy, and damp weather conditions in the NYC metropolitan area. The land-sea temperature differences in all seasons have a strong local impact on NYC climate. Land-sea breezes cool NYC on warm spring and summer days and warm NYC on cold nights in autumn, winter, and early spring (Gedzelman et al., 2003). Surface elevation increases from the coast to inland. In the NYC metropolitan region, two ridges are parallel to the -8 -coastline (Fig. 2a). One is along the boundary between Brooklyn and Queens, and the other is on the west side of the Hudson River. Each ridge blocks the cooler land-sea breeze when moving from the coast to inland, resulting in warmer surface temperatures (Fig. 2b).

Data source
Advanced Spaceborne Thermal Emission and Reflection Radiometer (ASTER) level 2 Surface Kinetic Temperature (AST_08, at 90 m spatial resolution) with an accuracy of 1.5 Kelvin was downloaded from NASA (ASTER_08, 2020, Yamaguchi et al., 1998). The temperature and emissivity separation algorithm accurately retrieves surface temperature using the ASTER multispectral thermal infrared (TIR) bands (Gillespie et al., 1998). QuickBird images with resolutions of 2.8 m were used to make the detailed land cover classification. In addition, parcel (tax lot) level urban structural information and the building footprint were downloaded from the NYC Department of City Planning (NYC DCP, 2020). The urban structural parameters include building footprint area, building classes, land use category, commercial area, residential area, total floors, building heights, and year built. According to the year built, more than 95% of the buildings were built before 2000. Tree census data of 2005 in NYC census tracts were also used in the analysis (NYC DEP, 2020).

Image classification
Urban areas represent a wide range of land uses and surface properties. To fully understand the spatial characteristics of urban structural and anthropogenic factors and the feedback to humandominated urban ecosystems, we classified the urban areas into different surface types. The most common subdivisions of urban areas are residential (low, medium, and high), industrial, commercial areas (Loveland and Belward, 1997). Each subdivision has distinct surface morphological characteristics defined by the amounts and types of vegetation and the size or roughness of different elements. These differences result in variations of flux partitioning across a city and the development of distinct micro-to local-scale climates. However, we used a slightly different classification scheme for different urban components based on spectral reflectivity of multispectral (blue, green, red, and near-infrared wavelengths) QuickBird satellite.
We used a supervised approach to classify the very high-resolution QuickBird imagery into various urban land categories (bright, medium bright, and dark impervious surfaces, shadows, trees, grassland, and bareground. We used the Fuzzy ARTMAP for classification, a neural network developed by Carpenter et al. (1992). This algorithm has been widely used for satellite image classification (Carpenter et al., 1999a,b;Pax-Lenney et al., 2001). The ARTMAP is a match-based learning neural network that uses a self-organizing arbitrary system to map inputs to outputs. It also has attractive features such as being fast and stable (Carpenter et al., 1999a,b).
We selected training areas for shadows, bright, medium bright, dark impervious surfaces, trees, grassland, and bareground. The spatial resolution of the QuickBird image is fine enough to distinguish these individual features represented by urban materials such as pavement, rooftops, trees, or bareground. The shadow category includes shadows cast by buildings and trees; bright impervious surfaces correspond to bright surfaces seen in the original satellite images, including most rooftops and industrial plants. Medium-bright surfaces are mainly concrete materials, and dark surfaces are mainly asphalt and tar. Dark surfaces appear brighter compared to shadows.
Trees and grasses are from parks and sides of streets, and bareground refers to playing fields.

Accuracy assessment
To quantitatively evaluate the accuracy of our image classification results, we used both confusion matrix and further visual interpretation. To create the confusion matrix, we recollected our ground truth from the original image through visual interpretation and our knowledge of New York City.
The ground truth used to create a confusion matrix is wholly independent of the training sites for our classification algorithm. In reality, confusion matrix analysis is not a panacea for classification accuracy assessment (Friedl et al., 2000). We collected large homogeneous areas as ground truth and avoided mixed or heterogeneous areas. The accuracy assessed using the confusion matrix tends to be overestimated. This problem is even worse for the urban area due to the small scale of the urban structure, resulting in many mixed pixels even for high-resolution QuickBird images.
Most of the pixels are correctly classified (Table 1). However, some grasslands are misclassified as medium-bright impervious surfaces, while some dark impervious surfaces are misclassified as medium bright surfaces or the shadow category. A few tree pixels are classified as grassland, a few medium bright impervious surface pixels are classified as bright surfaces, and a small part of bright surfaces are misclassified as medium bright surfaces. Overall the classification accuracy is 0.959, and the Kappa coefficient is 0.947.

Relationship between urban compositions and surface temperature
The urban land cover compositions derived from the QuickBird image were co-registered with the ASTER surface temperature. The fractional cover of QuickBird classes within each ASTER pixel was calculated. We assumed that a minimum of 10% pixel of each QuickBird class should substantially impact surface temperature in each ASTER pixel. The analysis revealed that surface temperature decreases with the increase in shadow coverage (Fig. 4a). The Sun position when the ASTER data was collected (52-degree Sun elevation angle) was lower than the Sun position when the QuickBird data was collected (63-degree Sun elevation angle), so the ASTER image has more shadows than the QuickBird image. As discussed below, this may not affect the relationship between surface temperature and shadows but may add noise to the relationships between surface temperature and impervious surface classes.
The surface temperature decreases with tree cover, with a steeper decrease than the shadow class (Fig. 4b). For the dominant grassland class, surface temperature shows no relationship (Fig.   4c). Grassland does not reduce and amplify urban heat, consistent with the ground observations conducted in a tropical city by Nichol and Wong (2005). Surfaces with brighter cover show an increasing surface temperature trend (Fig. 4d). Usually, bright surfaces with high albedo reflect more and absorb less solar radiation than other impervious surface materials. We expect that surface temperature decrease with bright surface cover. However, due to its low heat capacity, even the bright impervious surface can heat the surface easily and quickly. Also, the bright surface is surrounded by heat-trapping surfaces such as dark and medium-dark impervious surfaces. Those nearby darker materials may amplify and increase the surface temperature with bright covers.
Another contributing factor could be that most bright surfaces are rooftops and industrial plants, constantly exposed to the Sun. Therefore, the presence of co-existing darker impervious materials within the same ASTER pixel of 90 m spatial resolution.
The relationship is quite noisy for the dominant medium-dark impervious surface class, and the overall surface temperature increases with the medium-dark surface cover. However, the slope is not very steep (Fig. 4e). The main reason for the noise is that the definition for this class is not very clear, and many medium-dark surfaces can have different materials. Those materials can have different radiative emission and absorption properties that affect the surface temperature -16 -dramatically. The surface temperature increases with the impervious dark surface cover (Fig. 4f).
The slope is slightly steeper than the medium-dark surface cover.

Impact of land use and building structure on surface temperature
The mean surface temperature for building footprint in the Bronx is the highest, followed by Brooklyn, Manhattan, and Queens. Building footprint data indicate that the mean surface temperature in Staten Island is the lowest, at least 1.5 K less than that in Queens and 4 K less than that in Bronx and Brooklyn (Fig. 5). The analysis revealed that the industrial and manufacturing land use category generates significantly more surface heat than any other land use category (Fig.   6). In addition, transportation and utility, and parking facilities are also producing higher surface forms a high-temperature hotspot (Fig. 7). In parts of Bronx, Brooklyn, and Queens, multi-family walk-up buildings are also associated with higher surface temperature clusters. The possibility of anthropogenic factors such as winter heating and summer cooling is likely generating more heat in this locality.   The relationships between surface temperature and the number of trees in different NYC census tracts show that higher surface temperature is associated with lower tree counts, although some exception exists (Fig. 8a). Most of these census tracts are located in Bronx, Brooklyn, and Queens. However, low tree counts are also associated with lower surface temperatures in some of the census tracts. Those census tracts are primarily located in Manhattan. The scatter plot of the relationship between surface temperature and building height reveals an interesting pattern (Fig.   8b). Higher surface temperature is associated with lower building heights (i.e., fewer floors), whereas lower surface temperature is associated with higher heights (i.e., more floors). In addition to that, we observed that the surface temperatures increase with building density but show decreasing trend when density is higher than 0.4 (Fig. 9a). The scatter plot also revealed that surface temperature decreases with higher building height variability (Fig. 9b). Most of these census tracts that show a decrease in surface temperature are in Manhattan, where buildings are -20 -much taller with more floors and height variability. We also show that census tracts with less building height variability and higher tree counts are associated with lower surface temperature.
These results indicate that the larger variations in building heights and trees contribute to lower surface temperature in urban areas of New York City during autumn. The combination of the above factor increases shadows cast by higher-rise buildings on the surrounding lower-rise buildings causing cooling effects. The observed heterogeneity in surface temperature is likely associated with the neighborhood's specific building land use and structure. We observed that Manhattan has a greater share of high-rise buildings, while Staten Island has more trees than other localities. On the contrary, Bronx, Brooklyn, and Queens have a greater share of industrial and manufacturing land use and multi-family walk-up buildings, including fewer trees.

Discussion
Three factors, trees, building heights, and impervious surfaces, including bright surfaces, are primarily responsible for surface temperature heterogeneity in our study site. The replacement of vegetation by heat-trapping and non-porous urban materials alters surface conditions such as albedo, thermal capacity, and heat conductivity. Such transformation alters radiative fluxes between the surfaces and the lower atmosphere (Oke, 1982).
Trees reduce heat in two ways. Firstly, the shadows resulting from the tree control the total amount of radiation absorbed per unit surface area of heat-trapping materials. Secondly, trees return more surface heat to the atmosphere through evapotranspiration and reduce surface temperatures. Similarly, shadows cast by high-rise buildings reduce the amount of solar energy absorbed by the urban heat-trapping surfaces, and so there will be less heating effect. Guo et al. in New York City.
The future work will involve simulating time-series surface temperature patterns as a function of urban structures, vegetation cover, and shadows derived from high-resolution satellite data for the entire NYC region. It is well known that building shadows show strong seasonality and affect urban land surface temperature (Yu et al., 2019). Therefore, we planned to use high-resolution ECOSTRESS images to study the spatiotemporal pattern of urban heat. In particular, we will model the shadow effects due to building height variation and trees on surface temperature patterns. We will combine the locations of cool roofs and green roofs and quantify the cumulative effect on reducing the surface temperature in the city to help design priority locations since hightemperature hotspots during autumn are in the south Bronx, southeastern Queens, and northern Brooklyn. Our long-term goal is to assimilate satellite-observed spatiotemporal patterns of urban heat into urban climate models. The shadow analysis would provide helpful information in sustainable urban design, including access to sunlight during different seasons, and help mitigate the impact of climate change. The quantification of UHI effects has a potential public health implication and help safeguard people from heat-related stress.

Conclusion
Our study shows that the trees and shadows cast by high-rise buildings and their variability have a cooling effect. In contrast, more impervious surfaces show a heating effect even in the presence of highly reflective bright surfaces. The census tract with industrial and manufacturing areas and multi-family walk-up buildings as dominant land use categories correspond to the highest surface temperature. Buildings with lower heights (fewer floors) and height variability are associated with higher surface temperature. Although the building density is the highest in Manhattan (the central business district), many tall buildings with variable heights have shown cooling effects. Staten Island has the lowest mean surface temperature amongst all boroughs of New York City, where the number of trees is more. While, Bronx has the highest mean surface temperature and constitutes moderate building density, height, and height variability. The finding from this study has an important implication for urban heat island modeling, especially the positive effects of trees and