A Spatial Analysis of Urban Tree Canopy Using High-Resolution Land Cover Data for Chattanooga, Tennessee

Abstract

Urban tree canopy (UTC) provides urban residents with numerous benefits, including positive mental and physical health, the mitigation and prevention of urban heat islands, and a sense of place. Numerous studies have shown that as the wealth of a community decreases, so does the amount of UTC found in the community; thus, wealthier communities are more likely to enjoy the benefits that urban forests provide. Emerging technologies in remote sensing and GIS are allowing for new opportunities to study and understand the relationships between urban neighborhoods and UTC. In this study, land cover data for Chattanooga, Tennessee were derived from high-resolution (50 cm) multispectral imagery to assess the previously unknown extent and distribution of UTC and to measure the extent of UTC by neighborhood and census block group level. Using exploratory regression analysis, variables representing income, population density, race, educational attainment, and urban heat islands were analyzed to investigate the influence of UTC on neighborhood characteristics. This study found that UTC represented half of the total land cover composition, the tree equity was not as profound as shown in other cities, and the lack of UTC likely influences the prevalence of urban heat islands. This study also shows the importance and utility of using high-resolution imagery and land cover to assess and understand the impact and distribution of UTC in urban environments.

1. Introduction

Urban forests, which include all trees, shrubs, lawns, and other vegetation types, provide numerous ecosystem services to surrounding residents. Past studies have shown that urban forests, also referred to as urban tree canopy (UTC), are statistically associated with the economic health of city neighborhoods. A recent study of Atlanta, Georgia demonstrated the value of tracking UTC to inform city planning [1]. In the same region, a 2018 study conducted by Georgia Tech for Atlanta found that 47.9% of its land cover is represented by UTC and this number varies across the city due to zoning regulations, parks, and riparian areas [2]. UTC has been observed to provide ecosystem resilience, reduce the effects of urban heat islands, and serve as a carbon sink [3]. In contrast, neighborhoods with lower percentages of tree cover tend to have higher rates of poverty, lower educational attainment, lower home values, and higher populations at risk [4,5]. Impervious surfaces associated with the urban landscape may lead to increased stormwater runoff, affecting the water quality of streams and causing decreased evapotranspiration, increased solar radiation absorption, the additional release of anthropogenic heat, and changes in surface friction, resulting in changes to the near-surface air temperature, humidity, wind speeds, and precipitation; this effect is commonly referred to as a “heat island” [6]. As a city becomes larger and denser, the effect of these urban heat islands becomes more prevalent [7]. UTC is important in understanding a community’s tree resources and is most useful when overlaid with other geographic information, such as socioeconomic data, health data, and urban heat island locations [8]. This information helps planners and decision-makers to understand the distribution and effects of UTC for activities regarding goal setting, tree planting, park management, policy making, and others related to urban forests and greenspace.

A recent study shows that the City of Chattanooga, Tennessee has been urbanized significantly [9] and it behaves like an urban heat island (UHI) [10]. However, the relationship between UTC and socioeconomic variables and urban heat islands for Chattanooga, Tennessee has not otherwise been studied, at least to the authors’ knowledge. Chattanooga city officials and forest managers, along with local foundations and non-profit organizations, are interested in understanding this relationship and the impact of urban forests on local neighborhoods. The results of this study can help decision-makers to understand these relationships for Chattanooga and inform the development and implementation of an urban forestry plan that includes tree planting and management, neighborhood place making, park planning, and overall improvements in life in the urban landscape. These results may also benefit other cities, as an increased understanding of the relationship between urban forests and at-risk communities can create equitable neighborhoods.

1.1. High-Resolution Imagery and the Benefits of Urban Forests

Remote sensing is a well-established field and tool used in managing natural resources and in analyzing land cover change [11]. Since the launch of NASA’s integral Landsat program in 1972, Landsat and other sensors have collected millions of Earth images, providing valuable insights into a changing planet [12]. Most of these sensors collect images at a 10 to 30 m resolution, which allow for large-scale analysis but lack granularity when analyzing city neighborhood levels or finer resolutions. In recent years, widespread advances, predominantly led by commercial imagery vendors providing high-resolution imagery (1 m or finer) at higher temporal frequencies, have revolutionized landscape monitoring. For example, the Earth observation company Planet provides customers with analysis-ready multispectral imagery on demand. Their “PlanetScope” system maps the entire Earth daily, providing customers with 3-m-resolution multispectral imagery. Similarly, their “SkySat” sensors offer customers 50-cm imagery for any location on Earth at any time through custom tasking. Some of their notable customers include NASA and the Jane Goodall Institute.

The GIS software and data provider Esri recently announced an offering of 60-cm-resolution imagery to its customers at no additional cost, in partnership with the imagery vendor Maxar. This high-resolution imagery will be incorporated into their base maps as multispectral imagery that can be pulled directly into GIS software from the cloud [13]. Organizations and individuals can collect their own high-resolution imagery with inexpensive unmanned aerial vehicles (UAVs), and LiDAR data are also readily available.

Because of this wealth of affordable high-resolution imagery, GIS and remote sensing analysis techniques are adapting and evolving rapidly to unlock the vast geographic information that can be derived from these resources, often in near-real time, creating new opportunities and tools for researchers, natural resource managers, conservationists, and other decision-makers [14].

1.2. Study Area

Chattanooga, Tennessee (Figure 1) is a mid-sized city, 142.9 square miles in size, situated in Eastern Tennessee within the Southern Appalachian Mountains, between the Cumberland Plateau and the southern Blue Ridge Escarpments [15]. Chattanooga was once called “the most polluted city in the US” by Walter Cronkite in 1971 [16]. Since then, the city has invested heavily in addressing its pollution problems and is now a well-known destination for tourism and outdoor recreation. Despite this, Chattanooga was ranked by Climate Central, an independent group of scientists and communicators who research and report the facts about our changing climate, as the sixth fastest-warming city in the US, due to climate change [17]. Development is expected to increase along the major transportation corridors that connect Chattanooga to the nearby cities of Atlanta, Georgia and Nashville and Knoxville, Tennessee, mirroring the growth trends of the rest of the US [18].

An analysis by the University of Tennessee at Chattanooga revealed that, since 1984, Chattanooga has seen a 43% reduction in its urban forests and a 134% increase in developed land, based on collected Landsat imagery. The US Census Bureau reported the total population of Chattanooga in 2000 to be 155,600 people. For 2020, the US census reported a total population of 181,099, an increase of 25,499 people. The 2020 Census estimates a minority population of 79,514 people or 44% of the total population. The American Community Survey reported that, in 2020, the median household income was USD 47,165 per year and the median price of a home was USD 185,090; 31.4% of the population holds a bachelor’s degree or higher, with an unemployment rate of 5% [19]. This study hypothesizes that as UTC increases, so does the overall wealth of the neighborhood.

1.3. Mapping Urban Tree Canopy

To understand the correlations between urban tree canopy (UTC) and other variables, the spatial distribution of trees within a given geography must first be understood. The first step in performing a UTC assessment for a community is to define goals that align with their objectives and values. These goals should be set with the collaboration of community stakeholders and community needs and values related to issues such as storm water management, shading, equity, and public health. Having clear goals at the start of an analysis helps to determine the data needs and analysis methods; a common deliverable from such assessments is a suitability map identifying areas in need of tree planting projects or forest preservation [8,20,21].

There are numerous methods to measure UTC using geographic information. At smaller scales, detailed on-the-ground inventories can be conducted where field workers manually count, identify, and map trees. These types of manual approaches are not feasible for large extents, such as a city-wide analysis, given the mass number of trees to be counted, private landowner access issues, and limitations of time and other resources. Techniques using GIS and remote sensing are the most common and convenient tools in conducting UTC assessments [22].

One traditional approach to assessing UTC is using the “dot grid method”. A set of random points is cast upon a grid on either a printed or digital aerial image and the points are interpreted to determine tree cover by the center of each point. The presence or absence is then tallied for each point [23]. This approach is simple, and free digital GIS tools for this method, such as i-Tree Canopy, are available, but this technique is both time-consuming and often lacks a high spatial resolution [8].

Another common and economical method is to use GIS and open-source data provided by federal agencies to measure UTC for a given area. One common dataset, the National Land Cover Database, hosts land cover type, tree canopy cover, impervious surface, and land cover change maps at a 30 m resolution for all the United States. These data are provided by the Multi-Resolution Land Characteristics Consortium (MRLC), a conglomerate of federal agencies, every 5 years [24]. The advantages of the NLCD products are that they are free to the public and date back to 2001, offering the ability to easily perform spatial–temporal analyses of UTC. However, at a resolution of 30 m, the data are too coarse for neighborhood- and city-block-level analyses and tend to under-represent trees, and the temporal representation of the data might not align with a project’s needs and goals. Spatially, existing NLCD datasets would not be useful for analyses outside of the United States, although other options exist; for example, Esri’s newly released 10-m-resolution Land Cover Map of the World is derived from ESA Sentinel-2 imagery, with plans to update the map annually [25]. However, even a 10-m-resolution dataset still generalizes land cover at the neighbor and block scale.

Light Image Detection and Ranging (LiDAR) is another common method of mapping ground objects. LiDAR uses an active sensor attached to a plane, helicopter, or drone to scan an area with lasers to produce a dense point cloud representing the height attributes of objects such as trees, buildings, roads, parking lots, and bare earth across surfaces. LiDAR scans are highly accurate, often providing high resolutions of 2 m or better, and can penetrate cloud cover and shadows, which is problematic in other approaches to aerial image analysis. A drawback is the extreme cost of LiDAR and the infrequency of this type of data collection, which may mean that certain study areas have less coverage [8].

Multispectral imagery analysis is a common method of analyzing Earth objects to overcome the obstacles described above. In brief, light is reflected on the ground by objects, where certain wavelengths are absorbed and others are reflected. This gives different land cover types each a unique spectral signature. In a spectral image classification process, training samples are collected by an analyst and a machine learning algorithm is applied to identify other pixels in the image with similar spatial and spectral properties. These pixels are classified into land cover types by their pixel values [8].

Earth imaging satellites and other airborne sensors are usually sensitive to specific wavelengths of light, including near-infrared, invisible to the human eye. Multispectral sensors generate data ranging from the visible spectrum (red, blue, green) to near-infrared, while hyperspectral imagery collects hundreds of spectral bands. Perhaps the most common and suitable method of assessing UTC is utilizing high-resolution spectral aerial imagery and land classification techniques [26]. Land cover features in imagery are identified and extracted using automated techniques that produce spatially accurate and high-resolution custom land cover maps. A high resolution is often defined as 1 m or smaller, making this resolution more appropriate for neighborhood- and even parcel-level analysis [8].

When using high-resolution multispectral imagery, object-based image analysis (OBIA) classification offers accuracy advantages over traditional pixel-based classification techniques [27,28]. While traditional pixel-based classification considers only the spectral characteristics of the pixel clusters in multispectral imagery, object-based classification considers the spectral, geographic, geometric, and topological characteristics of the pixel clusters in multispectral imagery [29]. As a result, object-based classification approaches can yield more accurate, smoother, more interpretable land cover datasets compared to those of pixel-based approaches [30]. This segmentation and classification process can be time-consuming and requires large computational resources, especially when performed at the scale of larger municipality geographies [8].

Emerging techniques utilizing neural networks and deep learning are also proving to be an effective means to classify RGB imagery quickly and efficiently and identify objects such as trees and distinguish land cover types. Little research attention has been given to these methods in ecosystems science until recently, as these methods require a vast number of samples to feed into a deep learning algorithm and the results are often not conducive to field verification. Techniques using small subsets of LiDAR to perform unsupervised image classification have been developed to automatically collect samples. The results of the unsupervised classification are then improved with hand annotation and re-fed into the neural network to increase the accuracy [31]. Users and GIS data and software providers are sharing “deep learning packages” with pretrained files to apply to imagery. These packages usually require additional hand-annotated samples to derive meaningful results, as they are biased toward the geography and source imagery, yet they remain a viable option for the classification of trees by saving valuable time for spatial data scientists. These techniques show promise in utilizing the wealth of high-resolution imagery readily available from small satellites and drones, applicable to land management and environmental conservation.

For this study, an object-based image analysis and supervised classification method was used with high-resolution multispectral imagery, as described below in Section 2. Given the research needs and resources, it was determined that this was the most appropriate and accessible method for the classification of land cover and mapping of urban tree canopy for our study area.

1.4. Spatial Analysis

Regression analysis is a common method used in finding correlations and variables that explain, understand, model, or predict a certain phenomenon [32]. For this study, our dependent variable of UTC was tested for relationships with explanatory variables representing socioeconomic conditions and heat severity data representing heat islands, to help to answer the question, “is there a correlation between the percentage of UTC and income, race, education attainment, or other variables?” This is a common approach suggested in the literature when investigating the distributional equity of UTC [4].

2. Methodology

2.1. Object-Based Image Analysis

A high-resolution (50-cm) land cover map of Chattanooga was used and created by the University of Tennessee at Chattanooga IGTLab using object-based image analysis classification [33]. Stakeholders requested the most recent and highest-resolution multispectral imagery available and within the scope of the project budget. Funding from the Lyndhurst Foundation facilitated the tasking of high-spatial-resolution multispectral images of Chattanooga acquired from Planet’s SkySat constellation, a series of 21 high-resolution space-based Earth-imaging satellites in orbit around the Earth. The spatial resolution of SkySat is approximately 50 cm, meaning that every picture element, or “pixel” in a SkySat image, represents 50 cm on the Earth’s surface. Over the course of the late summer and early fall of 2021, Planet tasked its satellites to collect imagery on cloud-free days, supplying the research team with thousands of raw images in the form of 6 mosaic raster datasets to analyze and classify into land cover data.

During 2021 and early 2022, each of these 6 mosaics of orthorectified pansharpened imagery were segmented and classified into a thematic land cover map using object-based image analysis. In preparation for segmentation, the red, green, and blue bands were extracted from the imagery using the “Extract Bands Raster Function” within ArcGIS Pro. Segmentation was performed using the “Segment Mean Shift” geoprocessing tool after the optimal segmentation parameters were determined using the “Segment Mean Shift” raster function within ArcGIS Pro 3.2, Esri, USA, to save time and computational processing. Following imagery segmentation, training samples for image classification were collected using the “Training Samples Manger” within ArcGIS Pro 3.2, Esri, USA. Between 150 and 300 samples were collected for each land cover class, for each mosaic. Using these training samples, the segmented images were classified into a thematic land cover layer using the “support vector machine” supervised classification algorithm, as it is known to work well with an unbalanced number of classes and/or small sample sizes in each class. This resulted in 6 thematic land cover maps derived from the original imagery mosaics, which were combined into one land cover map of the city. This process was performed several times to achieve optimal results.

The resulting raster was found to have a 91% accuracy rate with a 0.9 kappa coefficient by sampling 500 random points across imagery and checking for their correct land cover classification (

Table 1. Confusion matrix derived from the accuracy assessment on the refined land cover dataset.

Class Value	Forest	Non Forest Veg	Impervious	Exposed Soil	Water	Total	User Accuracy	Kappa Coefficient
Forest	232	12	3	0	0	247	0.939271	0
Non Forest Veg	8	90	1	2	0	101	0.891089	0
Impervious	0	3	117	1	0	121	0.966942	0
Exposed Soil	0	0	2	8	0	10	0.8	0
Water	1	0	0	0	27	28	0.964286	0
Total	241	105	123	11	27	507	0	0
Produce Accuracy	0.96266	0.857143	0.95122	0.727273	1	0	0.934911	0
Kappa	0	0	0	0	0	0	0	0.902254

) The resulting land classes included forests, non-forest vegetation, developed areas, water, and bare earth (Figure 2 and Figure 3). These generalized classes were chosen to assess major land cover groups within the project duration, as a simplified alternative to the NLCD classes. This thematic 2021 land cover map of Chattanooga is the highest-resolution and most accurate map ever produced for the city.

2.2. Census Block Groups (CBG)

Census block groups (CBGs) were used to aggregate the total area of canopy cover (UTC) data and socioeconomic variables. CBGs are statistical division areas used by the US Census Bureau that closely resemble the sizes of neighborhoods containing between 600 and 3000 people [19]. The neighborhood boundaries for Chattanooga are arbitrary and often contested. A GIS layer of CBGs was obtained from the US Census TIGER/Line Files (Topologically Integrated Geographic Encoding and Referencing) program and clipped to the city boundary of Chattanooga. Demographic data were obtained from the American Community Survey and Esri’s demographic data program (

Table 2. Variables used for this study.

Attribute	Source
Mean Household Income	American Community Survey (2020)
Population Density	American Community Survey (2020)
Educational Attainment: Bachelor’s Degree	American Community Survey (2020)
Race White	American Community Survey (2020)
Race Black	American Community Survey (2020)
Race Hispanic	American Community Survey (2020)
Heat Severity (Heat Islands)	Trust for Public Land (2021)

). Heat island data were sourced from the Trust for Public Land Heat Severity Dataset. These data variables were aggregated to the CBGs along with the total area and percentages of land cover types derived from the thematic land cover layer using the ArcGIS Pro geoprocessing tools “Summarize Within” and “Data Enrichment” (Figure 4 and Figure 5). The final layer resulted in a dataset containing CBGs with their unique FIPS codes, socioeconomic variables for population demographic variables, a heat severity index for the identification of heat islands, and percent canopy cover area for statistical and spatial analysis (Table 2). These variables, described in Table 2, were selected from a combination of past studies identified in the literature [34] and stakeholder input from the City of Chattanooga and the nonprofit Green|Spaces, which focuses on sustainability and equity issues for Chattanooga neighborhoods.

3. Spatial Analysis

3.1. Land Cover Statistics

The resulting thematic land cover types derived from the Planet SkySat imagery were converted in a vector data layer and used to summarize the total area and percentages of land cover classes for the city boundary of Chattanooga and approximate neighborhood boundaries within the Chattanooga city limits. The most dominant land cover type for Chattanooga is forested land cover, at 49% (Figure 6). The mean forested area for neighborhoods is 45%, with the Downtown, Highland Park, and Glenwood neighborhoods having the smallest amount of UTC in the city, at less than 30% forested land (Figure 7).

A uniform percent canopy cover layer was also created by generating a hexagonal grid for the city, with hexagons totaling 100 acres in size, using the ArcGIS Pro geoprocessing tool “Summarize Within”. The total area of forests was calculated for each grid and displayed as a percentage (Figure 8), giving an equal comparison of UTC density across the city [35].

3.2. Exploratory Regression

Regression analysis is a standard statistical method for the prediction or explanation of relationships between variables and their dependents [3,32]. The explanatory variables listed in column one of

Table 3. Exploratory regression results.

Explanatory Variable	Adjusted R2	p-Value	Correlation
Urban Heat Islands (2021)	0.31	0.01	negative
Population Density (2020)	0.02	0.05	negative
Mean Income (2020)	0	0	positive
% Bachelor’s Degrees (2020)	0	not significant	positive
Race: % White (2020)	0.02	0.1	positive
Race: % Black (2020)	0.02	0.05	negative
Race: % Hispanic (2020)	0.05	0.05	negative

were used in a bivariate exploratory regression analysis to investigate the statistical relationships between demographics and heat islands at the census block group level.

The results of the bivariate exploratory regression analysis show associations with these variables and the percentage UTC; however, the statistical relationships overall are weak due to low p-values in many cases. This indicates that even though there is a correlation, educational attainment, race, and income likely do not predict or influence UTC for Chattanooga, but they are associated with the percentage of UTC. Heat islands/heat severity and population density were the strongest predictors of lower percentages of UTC for Chattanooga; as the percentage of UTC decreases, the population and heat severity will also likely increase, but the adjusted R² for both is still low overall. By comparing the distribution of the percentage canopy from Figure 8 to the distribution of the heat severity in Figure 9, it is found that areas that are low in UTC tend to be more susceptible to heat severity, including areas such as Downtown, which has a high rate of impervious surface, as shown in Figure 2. The lack of strong correlations between income and other dependent variables and UTC is likely attributed to the high rate of UTC for Chattanooga, with forest cover representing half of the land cover for the city.

4. Discussion

This study shows the need and utility of high-resolution imagery and object-based image analysis to quantify land cover for city neighborhood-level analysis. High-resolution land cover data are essential to understand the relationships between the percentage urban tree canopy cover, demographics, and urban heat islands. These data reveal insights into which neighborhoods have adequate urban forests and where efforts by city planners should be made to manage, restore, and preserve them. The 30-m resolution National Land Cover Dataset for 2019 shows forested land cover for Chattanooga, where = 27% of land cover types, versus 49%, using 50-cm-resolution land cover data derived from Planet SkySat imagery. This confirms that coarser and more available land cover products underrepresent UTC in urban environments. The monetary costs of acquiring and processing high-resolution imagery and land cover data to understand UTC should be considered a worthy investment for urban planners that wish to have a full understanding of their urban forests and how they relate to neighborhood residents. One limitation of this study was the availability of high-resolution heat severity data to complement the 50-cm-resolution land cover data used. The Trust for Public Land’s heat severity datasets are derived from Landsat at a 30 m resolution, possibly generalizing to heat severity neighborhood-level analysis.

From the initial exploratory regression analysis, there is some evidence supporting the hypothesis that income and other demographic variables for Chattanooga are correlated with UTC, as statistical relationships were found; however, the coefficient of determination values illustrated in Figure 10 are too low to draw any definitive conclusions. Still, many other situational factors may also explain these correlations. These low adjusted R-squared values suggest that there is little evidence that these variables predict or influence the amount of UTC. In this local case, it might be that the exceptionally high percentage of urban canopy in Chattanooga skews the results. Additional studies of other urban areas could be conducted to see if overall cities with smaller urban forest percentages show a stronger correlation with the other socioeconomic indicators. More research exploring the variables included in this study, and others, should be conducted from a temporal and spatial perspective to better understand the influence, or lack thereof, between social demographics, environmental characteristics, and UTC for Chattanooga. This study does show, however, some evidence that as the percentage of UTC decreases, the heat severity and heat island effects increase. However, here, too, more research in this area should be conducted in the future to fully understand the influence of UTC and heat islands. It should be noted that trees still bring numerous benefits to people and urban ecosystems, and city planning policies should aim to maintain as much of the tree canopy as possible, regardless of these results. If nothing else, avoiding the expansion of urban heat islands by increasing urban forests aids community resilience to climate change. The lack of relationships among these socioeconomic variables indicates that the UTC coverage is somewhat uniform across neighborhoods and the city should strive to maintain its existing urban forests and increase them where the percentage is low.

The distribution and relationships between UTC and people are inherently a geographic problem, and GIS is essential in understanding the relationships and distribution of the two. The literature suggests that UTC provides numerous benefits to urban communities and that GIS, remote sensing, and spatial analyses are valuable and essential tools for communities to understand the distribution of UTC and the impacts that it has on at-risk populations. Spatial analyses, such as those performed in this study, can be used to inform policy and used in the decision-making process for the management and improvement of UTC for all urban communities. As cities grow over time, the effects of climate change increase as the natural land cover is converted; thus, it is vital to understand the associations between UTC and its benefits to the urban environment. Decision-makers may be more acutely aware of the detriments to any populations lacking UTC benefits and adjust their policies accordingly. Therefore, these types of spatial analyses should be performed routinely to understand gaps in tree cover and where these gaps need to be filled in order to maintain sustainable and build more equitable communities.

5. Conclusions

Chattanooga is quite literally a “green city”, with roughly half of its area composed of forested land, as shown in our 50-cm-resolution land cover map of the city—the highest-resolution and most accurate land cover map that has been created for Chattanooga, according to the authors’ knowledge. Historical Landsat imagery analyses indicate a reduction in UTC with new development. Studies such as ours need to be performed at defined intervals to track trends over time and to monitor urban forested land. Precision GIS and remote sensing are crucial to understanding the relationships between land cover and human/social interactions. Even though strong correlations between UTC and income/other demographics were not found, it is still important to protect and enhance the existing UTC for biodiversity, health, and a sense of place and to help to combat heat island effects and promote climate resilience. The spatial analysis described in this paper should be conducted regularly to track changes within UTC and other land cover variables, heat severity, and demographics to minimize the detrimental impacts on urban forests, thus preventing the emergence of tree equity and access issues for residents. This research shows the need for future studies of UTC and its influence on social and environmental variables. Future research could help us to better understand the topics described in this paper and identify temporal trends.

The preservation of UTC depends on various stakeholders’ interest and values throughout the city. Future policies, incentives, and outreach all stand to benefit from a better understanding of the advantages of protecting and expanding UTC for all residents of Earth beyond Chattanooga. Partnerships between governments, local non-profits, and private citizens and industry will help these management goals to become a greater possibility. As Chattanooga has maintained a track record of success in becoming more environmentally conscious, earning the title of “The Scenic City” and striving to meet sustainability goals, continued UTC monitoring is imperative in understanding the environmental health of the city and preserving its qualities.