Spatiotemporal Analysis of Urban Forest in Chattanooga, Tennessee from 1984 to 2021 Using Landsat Satellite Imagery

Abstract

Chattanooga, Tennessee is one of many cities in the Southeastern United States that is experiencing rapid urban growth. As these metropolitan areas continue to grow larger, more and more of Earth’s unique temperate forest, an ecosystem of enormous cultural, ecological, and recreational significance in the Southeastern United States, is destroyed to make way for new urban development. This research takes advantage of the extensive temporal archive of multispectral satellite imagery provided by the Landsat program to conduct a 37-year analysis of urban forest canopy cover across the City of Chattanooga. A time series of seven Landsat 5 scenes and three Landsat 8 scenes were acquired between 1984 and 2021 at an interval of five years or less. Each multispectral image was processed digitally and classified into a four-class thematic raster using a supervised hybrid classification scheme with a support vector machine (SVM) algorithm. The obtained results showed a loss of up to 43% of urban forest canopy and a gain of up to 134% urban land area in the city. Analyzing the multidecade spatiotemporal forest canopy in a rapidly expanding metropolitan center, such as Chattanooga, could help direct sustainable development efforts towards areas urbanizing at an above-average rate.

1. Introduction

Humans have directly impacted the health and integrity of temperate forests more than any other forest type on Earth [1]. The current global extent of temperate forests is now estimated to be 40–50% of its historic expanse [2]. By as early as 1100 BC, temperate forest areas in Europe had been reduced to 30% of its original extent because of increased demand for fuel wood and agricultural land area by early human civilizations [1,3,4]. During the American Industrial Revolution, the rapid clearing of forests for timber harvest and agricultural land conversion cleared 35% of the continent’s temperate forests, leaving a mere 65% intact [4].

Today, temperate forests cover about 10.4 million km², representing a quarter of the world’s forest cover today. North America contains roughly 60% of total global temperate forests, Europe contains 24%, and Eastern Asia, combined with a handful of other areas in the Southern Hemisphere, contains the remaining 16% [4]. Despite the overexploitation of temperate forests to fuel growing timber markets during the pre- and post-industrialization stages of emergent nations, temperate forest land area is slowly increasing as successional communities begin to reestablish themselves in areas historically used for timber harvest or agricultural production that are no longer in operation [4,5]. However, in the same way the overexploitation of timber in the industrial revolution threatened forest habitat and its respective carbon reservoir in the 19th century, rapid urban development and the associated habitat loss is threatening the global extent of temperate forests and its carbon reservoir once again, now at an unprecedented rate, scale, and intensity.

1.1. Global Forests and Carbon Storage

Today, 30% of Earth’s surface is covered by forest [6]. In addition to the myriad human and animal communities, their structure, and ecological function support, the trees of Earth’s forests represent a staggering 80% of the total biomass on Earth [5,7,8]. In fact, the global forest carbon reservoir alone has incorporated more atmospheric carbon in its biomass and soils than currently exists in the entire atmosphere [5]. Once sequestered [9,10], the carbon making up the woody biomass of trees can remain stored until the end of that tree’s life cycle, which can range anywhere between 50 and 5000 years depending on the species of tree and the geographic region in which it grows. Therefore, the trees of Earth’s forests represent a prodigious reservoir in the global carbon cycle [11,12,13].

1.2. Urban Forests, Their Significance, and Future

As urban centers in the United States, Europe, and Eastern Asia grow ever larger to accommodate swelling populations, temperate forests are cut, bare earth is paved, and buildings are constructed. This generalized process of land cover change is collectively referred to as urbanization [14,15]. Due to their proximity to human infrastructure and population centers, urban forests, or forested areas within or adjacent to a metropolitan center, are often the best option for developers [16,17]. The urban forest of a given city includes forested fragments, greenways, riparian zones, wetlands, urban parks, residential trees, street trees, and working forests [18]. The US Forest Service reports that approximately 127 million acres of forest in the US is in immediate proximity to metropolitan areas and, therefore, can be classified as urban forest [19]. However as stated previously, when forest habitat is closer in proximity to human development, it runs a higher risk of being negatively impacted or destroyed [20,21]. Therefore, due to the collective burden of natural disturbances (forest fires and extreme weather events) and anthropogenic pressures (losses in habitat and habitat connectivity resulting from land cover changes, associated decreases in biodiversity, invasive species proliferation, and any unanticipated indirect and/or synergistic reactions between the aforementioned pressures), the composition, structure, and function of urban forests are at an extreme risk of deterioration.

Urban forests are critically relied upon by wildlife, as following the urbanization of a landscape, the remaining interspersed fragments of urban forest serve as functional islands which provides extant species with some level of shelter from human impacts [22]. Urban forests provide several essential services to humans as well. Urban forests play a significant role in the establishment of one’s sense of place, which can be passed down through generations [23]. Therefore, forests often possess great cultural value. Additionally, urban forests work to mitigate a potential urban heat island effect by cooling surface and air temperatures via evapotranspiration, reduce the volume and rate of flow of a runoff event via stormwater uptake, filter city air, and reduce urban noise [24].

Contrary to what has been observed in tropical forests, recent research is suggesting that the trees of temperate forest fragments adjacent to urban areas are sequestering carbon dioxide in biomass and soils at an accelerated rate [25,26,27]. A previous study conducted by Morreale et al. [26] found that trees along the edges of temperate forest fragments adjacent to urban areas grow up to 36% faster and sequester 24% more carbon than trees within the forest interior. Furthermore, another previous study conducted by Garvey et al. [27] found that in urban areas, respiration rates and associate carbon loss rates of soil along the edges of urban forest fragments are up to 25% lower compared to trees within the forest interior. Because of their enhanced ability to sequester carbon, urban temperate forests represent a critical sink in the global carbon cycle.

As metropolitan areas grow through time, it is increasingly important to know exactly where a city’s urban forest areas are to monitor and mitigate negative impacts to their structure, function, and composition resulting from the regular exposure to the collective disturbance of the surrounding urbanized landscape.

1.3. Assessing Urban Forest Extent

1.3.1. Traditional Assessment of Urban Forest Extent

Traditionally, the identification of the extent of a city’s urban forest has been conducted by extensive and strategic field sampling across a representative sample of the city. However, due to the considerable amount of time needed to visit and document a statistically viable sample of urban forest sites across an entire city and the steep financial cost of hiring a third-party field surveyor, traditional field-based assessments of urban forest extent are less common.

1.3.2. Advantages of Remote Sensing Technology

In recent decades, the utilization of remote sensing technology alongside a geographic information system (GIS) has proven to be a powerful and cost-efficient tool with a wide variety of applications including land use and land cover change studies, meteorological studies, emergency response, water quality monitoring, and the monitoring of urban forest vegetation [28,29,30,31,32,33]. Remote sensing can be defined as the utilization of the electromagnetic radiation of objects on Earth’s surface based off a given object’s interaction with visible, infrared, and microwave portions of the electromagnetic spectrum. Objects selectively absorb and reflect electromagnetic energy due to differences in the molecular composition of their surface [34]. Functionally, remote sensors either detect electromagnetic radiation (EMR) from the sun reflected off objects on Earth’s surface (passive sensors), or they detect their own emitted EMR reflected back from objects on Earth’s surface (active sensors). Many remote sensors typically run continuously, collecting data along specific orbital paths, creating large volumes of reliable data in a short amount of time [34]. Additionally, remote sensors provide a synoptic view of Earth, allowing researchers to obtain data virtually anywhere on the Earth’s surface without the need to physically visit a specific location in the field [35].

1.3.3. Landsat: A Legacy of Earth Systems Monitoring

Of the host of remote sensors used to monitor forest vegetation, no other remote sensor has been as influential as those employed by the Landsat program [31,33,36,37,38,39,40,41,42]. Landsat’s temporal data archive provides continuous data from 1972 to the current day, making it the single earliest and longest continuous archive of global multispectral remote imagery [43,44]. Finally, one of the inherent reasons that Landsat imagery has been routinely used in the monitoring of vegetation is that, since 2008, researchers can freely view and download raw and processed data from Landsat’s extensive temporal archive.

1.4. Problem Statement and Research Objectives

The City of Chattanooga, TN is among many metropolitan centers across the United States that is experiencing rapid urbanization. Research conducted by Hall and Hossain [45] has confirmed that Chattanooga’s urban land areas have increased rapidly since 1986. Previous research has also confirmed that the conversion of forest to developed areas in the City of Chattanooga can be directly associated with impacts to surface water quality, increased surface and air temperatures, and decreased canopy cover resulting from the extraction of a landscape’s vegetation [15,45]. Additionally, temperate forests that are cut down to make way for urban development in the City of Chattanooga are not likely to be reestablished over time, thereby creating a permanent imbalance in the carbon cycle. Once extracted, all the carbon that the forested area sequestered in its biomass during its life cycle is released back to the atmosphere as carbon dioxide (CO₂), thereby reducing the total carbon sequestration potential of the urbanized land area and increasing the carbon footprint of the City of Chattanooga [8].

In order to help conserve the extant temperate forest habitat within and surrounding Chattanooga, and to mitigate the levels of CO₂ released to the atmosphere associated with the loss of temperate forest land area following urbanization, this research was designed to detect and map the historic extent of Chattanooga’s urban forest canopy along a specific time interval using multispectral imagery to facilitate sustainable development in rapidly urbanizing locations. This research took advantage of the extensive temporal archive of multispectral satellite imagery provided by the Landsat program to conduct a 37-year land cover change analysis across Chattanooga, Tennessee. A time series of seven Landsat 5 Thematic Mapper (TM) scenes and three Landsat 8 Operational land Imager (OLI) scenes acquired over Chattanooga, TN between 1984 and 2021 at an interval of about five years was obtained and analyzed for this study. Image processing and analyses were carried out using a supervised hybrid classification scheme with a support vector machine (SVM) algorithm.

To date, no published research has mapped the current and historic extent of Chattanooga’s urban forest. Therefore, this research bears incredible regional significance for science and development. It also presents a unique application of Landsat satellite imagery to study urban forests in a mid-size city of Southeast Tennessee.

2. Materials and Methods

2.1. Study Site and Data Collection

The study site boundary of this research includes the entire City of Chattanooga, Tennessee as shown in Figure 1. For this study, Landsat 5 Thematic Mapper (TM) and Landsat 8 Operational Land Imager (OLI) data were obtained from USGS’s Earth Explorer data hub, seen in Figure 2. Landsat data can be obtained from Earth Explorer as a Level-1 or a Level-2 product, where Level-1 data are provided as digital numbers without atmospheric corrections conducted, and Level-2 data are provided as calibrated surface reflectance values. For this research, Level-2 data were obtained. Downloaded from Earth Explorer as Level-2 products, Landsat 8 OLI imagery consisted of 9 bands, while Landsat 5 TM imagery consisted of 7 bands, as shown in

Table 1. Landsat 5 TM and Landsat 8 OLI sensor specifications.

	Bands	Description	Wavelength (μm)	Spatial Resolution (m)	Temporal Resolution
Landsat 5	TM 1	Blue	0.45–0.52	30	16 Days
	TM 2	Green	0.52–0.60	30
	TM 3	Red	0.63–0.69	30
	TM 4	Near-Infrared	0.76–0.90	30
	TM 5	Near-Infrared	1.55–1.75	30
	TM 6	Thermal-Infrared	10.40–12.50	120
	TM 7	Mid-Infrared	2.08–2.35	30
Landsat 8	OLI 1	Coastal Aerosol	0.43–0.45	30	16 Days
	OLI 2	Blue	0.45–0.51	30
	OLI 3	Green	0.53–0.59	30
	OLI 4	Red	0.64–0.67	30
	OLI 5	Near-Infrared	0.85–0.88	30
	OLI 6	Shortwave-Infrared	1.57–1.65	30
	OLI 7	Shortwave-Infrared	2.11–2.29	30
	OLI 8	Panchromatic	0.50–0.68	15
	OLI 9	Cirrus	1.36–1.38	30

.

In order to understand how forest canopy coverage was impacted by urbanization between 1984 and 2021, imagery was obtained along a defined interval, where the years between imagery acquisition (referred to in

Table 2. List of the acquired Landsat 5 TM and Landsat 8 OLI imagery.

Scene ID	Year	Month and Day	Acquisition Gap	Satellite–Sensor	RGB
1	1984	June 27	0	Landsat 5—TM	321
2	1988	July 8	4		321
3	1990	June 28	2		321
4	1995	July 12	5		321
5	2000	June 23	5		321
6	2004	July 20	4		321
7	2009	June 16	5		321
8	2014	June 14	5	Landsat 8—OLI	432
9	2019	August 31	6		432
10	2021	July 3	1		432

as ‘Acquisition Gap’) was equal to or less than 5 years. A 5-year interval was selected by the researchers to effectively observe urban growth and canopy loss through time without obtaining an excessive number of Landsat scenes. In this research a total of 10 scenes were obtained representing 37 years of urbanization in the City of Chattanooga, TN. All Landsat scenes were obtained during the months of June and July to account for any seasonal variation in canopy coverage except the one acquired in 2019.

In some cases when cloud coverage was an issue, imagery was acquired before the defined 5-year interval. In one case, however, the acquisition period between scenes was greater than the defined 5-year interval. This was due to excessive cloud coverage above Chattanooga, TN between the months of June and July from 2014 to 2020. For this reason, the acquisition gap between scenes 8 and 9 was 6 years instead of 5 years.

The flowchart in Figure 3 summarizes the workflows adopted in the methodology of this study.

2.2. Data Processing

The acquired image time series was processed using digital image processing (DIP) techniques [46]. The processing was performed using the DIP tools available on ERDAS Imagine 2020 and ArcGIS Pro 3.1 software. The images were acquired and processed as part of a master’s thesis research conducted by William Stuart [47]. The specific image processing steps performed are explained below.

2.2.1. Image Pre-Processing and Enhancement

Using ERDAS Imagine’s layer stack tool, the individual bands for each of the downloaded scenes, as shown in Table 1 and Table 2, were stacked to create 10 composite multispectral images. Next, using ArcGIS Pro’s raster geoprocessing tool, each of the 10 composite images were clipped to the City of Chattanooga’s boundary (Figure 1). From each of the 10 scenes, a true color image was derived using ArcGIS Pro. For Landsat 5 scenes 1–7, the band combination for RGB true color was 321. For Landsat 8 scenes 8–10, the band combination for RGB was 432. The images were then stretched using either percent clip or standard deviation depending on the clarity of the image after applying a stretch. Figure 2 shows the time series of Landsat 5 TM and Landsat 8 OLI images (from 1984 to 2021) used for this study. The images are shown in true color.

2.2.2. Feature Extraction

The next step in digital image processing was classifying each image to obtain a thematic land cover map in the following classes: developed areas, forest canopy, non-forest vegetation, and water. To accomplish this, both pixel-based supervised classification and pixel-based unsupervised classification were utilized. This is commonly referred to as a supervised hybrid classification strategy. For reasons related to Landsat’s 30 m spatial resolution, pixel-based supervised hybrid classification was ultimately selected.

2.2.3. Supervised Classification

The image classification wizard available on ArcGIS Pro was used to conduct supervised classification [46,47,48]. The classification was realized using the pixel-based approach relying on the training samples derived from the image. Figure 4 shows examples of the training sample polygons shown on the Landsat 8 OLI image acquired in 2021. The pixels of the images were classified into three classes: pervious surfaces, impervious surfaces, and water. The support vector machine algorithm was used as the machine learning classifier for supervised classification.

Satellite imagery acquired through the Landsat program has been widely used for land use and land cover mapping since 1972 [49,50]. The supervised classification technique using the Maximum Likelihood Classification (MLC) algorithm is the most popular and conventional method to perform a land use and land cover classification task with an acceptable rate of accuracy. The MLC classification is based on a parametric approach which implies the assumption of the selected signature classes within the normal distribution [51], which is not certain about the current study. Recent studies show that some non-parametric based classification techniques have been used for extracting major classification as well as sub-classification with better accuracy [52]. The widely used non-parametric classification techniques are Decision Trees, Fuzzy C-Mean, Artificial Neural Networks (ANN), and Support Vector Machines (SVMs). Among them, SVMs have been reported to classify satellite imagery to generate land use and land cover maps with better accuracy in comparison to MLC classification [53,54]. That is why this research used an SVM as the algorithm for supervised classification.

An SVM is based on statistical theory and is used for classification and regression problems [55,56]. It is defined as a machine learning algorithm that uses supervised learning models to solve complex classification tasks by performing optimal data transformations that determine boundaries between data points based on predefined classes. The primary objective behind SVMs is to transform the input data into a higher-dimensional feature space (hyperplane). This transformation makes it easier to find a linear separation between classes in the imagery. It is considered one of the best supervised classification algorithms because of its capability to handle high-dimensional data. It is very effective in cases with limited training samples.

The output data obtained from supervised classification were 10 raster datasets each consisting of 3 land cover classes: pervious surfaces, impervious surfaces, and water. Figure 5 shows the result of the supervised classification obtained for the 2021 image.

2.2.4. Unsupervised Classification

Using ArcGIS Pro’s geoprocessing tool (Extract by Mask), pixels from the original multispectral Landsat images coincident with the pervious surface class from the output of the previous step were extracted. In this step, unsupervised pixel-based classification [46,47,48] within ArcGIS Pro’s image classification wizard was employed to classify the extracted multispectral pervious pixels into 10 spectrally unique classes. An ISODATA clustering [46,47,48] algorithm was used as the classifier for unsupervised classification. The maximum number of classes was set to 10. The output data obtained from unsupervised classification were 10 raster datasets, each consisting of 10 spectrally distinct classes of pervious pixels. Figure 6 shows the result of the unsupervised classification for the 2021 image.

2.2.5. Post Processing—Image Reclassification

The unsupervised classification outputs needed to be reclassified into 2-class rasters consisting of forest canopy or non-forest vegetation. This was accomplished using a true color reference to inspect each of the 10 classes of pervious pixels for all unsupervised outputs. After determining whether each class of the unsupervised outputs belonged to the forest canopy class or the non-forest vegetation class, the unsupervised output was reclassified using ArcGIS Pro’s geoprocessing tool (reclassify). The output obtained from the reclassification of unsupervised data was 10 raster datasets, each consisting of forest canopy and non-forest vegetation. Figure 7 shows the result obtained after reclassification for the 2021 image.

Figure 6. Land cover output consisting of 10 spectrally distinct classes of pervious pixels (following unsupervised classification) for the 2021 Landsat 8 OLI imagery. Each color indicates a distinct feature within the extracted pervious surfaces.

Figure 6. Land cover output consisting of 10 spectrally distinct classes of pervious pixels (following unsupervised classification) for the 2021 Landsat 8 OLI imagery. Each color indicates a distinct feature within the extracted pervious surfaces.

Figure 7. Reclassification of 10 spectrally distinct classes of previous pixels to a 2-class thematic raster consisting of non-forest vegetation and forest vegetation derived from 2021 Landsat 8 OLI imagery.

Figure 7. Reclassification of 10 spectrally distinct classes of previous pixels to a 2-class thematic raster consisting of non-forest vegetation and forest vegetation derived from 2021 Landsat 8 OLI imagery.

2.3. Data Analysis

Finally, to obtain the desired 4 class land cover map, the outputs from the reclassification of pervious pixels and supervised classification needed to be combined. From each of the supervised classification outputs, all pervious pixels were reclassified to ‘no data’ using ArcGIS Pro’s geoprocessing tool (reclassify). Next, using ArcGIS Pro’s raster function (merge), the developed and water classes from supervised classification and the forest and non-forest vegetation classes from unsupervised classification were combined. The final output data obtained were 10 raster datasets, each consisting of 4 classes: impervious surfaces, urban forest canopy, non-forest vegetation, and water (

Table 3. Final thematic land cover dataset class names and descriptions.

Land Class	Class Code	Class Description
Impervious Surfaces	1	Buildings, Roads, Cars, Parking Lots, Artificial Turf, etc.
Non-Forest Vegetation	2	Grasses, Scrubs, Shrubs, Crops, Ornamental Plants, etc.
Urban Forest Canopy	3	All tree canopy within the study area.
Water	4	Flooded Wetlands, Rivers, Streams, Man-Made Retention Ponds, etc.

). Figure 8, Figure 9, Figure 10, Figure 11, Figure 12, Figure 13, Figure 14, Figure 15, Figure 16 and Figure 17 show the final output for all 10 images.

2.4. Accuracy Assessment

Typically, accuracy assessments of spatial analyses incorporate ground-truth data to reference the derived land cover dataset. However, this is not possible when dealing with temporal datasets. Therefore, Google Earth’s historic imagery archive was utilized as a proxy for ground-truth reference data. For each year of acquisition for the Landsat imagery obtained in this research, polygons representing pure samples of the 4 classes (Table 3) derived in the final output of this analysis were collected using random stratified sampling across Google Earth’s historic imagery and saved as separate KML files (Figure 18). The files were then uploaded to ArcGIS Pro and saved as feature classes. Finally, using ArcGIS Pro’s geoprocessing tool, confusion matrices were prepared, and an analysis of producer accuracy, user accuracy, and overall accuracy [57,58,59] was performed for each year. The user’s accuracy is calculated by dividing the total number of classified pixels that match the reference pixels by the total number of classified pixels for that class. The producer’s accuracy is calculated by dividing the total number of classified pixels that match the reference pixels by the total number of reference pixels for that class. Overall accuracy is the probability that an individual feature will be correctly classified with a classification algorithm. It is calculated by dividing the sum of the match pixels of all classes (with corresponding reference pixels) by the total number of pixels classified [57,58,59].

3. Results and Analysis

3.1. Quantification of Land Cover Classes

Areas and relative percentages of land cover classes were quantified for the 10 final land cover datasets. The results are shown in

Table 4. Quantification of land cover classes. Non-Forest Vegetation is symbolized by NF.

	Impervious Surface		NF Vegetation		Forest Canopy		Water
	Acres	% Area	Acres	% Area	Acres	% Area	Acres	% Area
1984	15,526	16.10%	20,229	20.97%	55,549	57.60%	5140	5.33%
1988	24,546	25.45%	18,595	19.28%	48,325	50.11%	4979	5.16%
1990	23,512	24.38%	22,976	23.82%	44,968	46.63%	4988	5.17%
1995	25,875	26.83%	23,931	24.81%	41,707	43.24%	4931	5.11%
2000	21,833	22.64%	26,213	27.18%	43,356	44.95%	5043	5.23%
2004	22,358	23.18%	30,393	31.51%	38,630	40.05%	5063	5.25%
2009	27,485	28.50%	28,235	29.28%	35,590	36.90%	5134	5.32%
2014	32,365	33.56%	24,086	24.97%	34,857	36.14%	5136	5.33%
2019	31,981	33.16%	22,188	23.01%	37,278	38.65%	4991	5.18%
2021	36,316	37.65%	23,347	24.21%	31,924	33.10%	4857	5.04%

.

3.2. Assessment of Spatiotemporal Trends

Figure 19 displays the area of each land cover class in acres for all ten images in a bar chart. The area in acres and the relative percentage of each land cover class for all images is also provided in Table 4. Finally, the percent change in land cover classes between consecutive imagery dates was calculated and provided in

Table 5. Land cover class area percent change between imagery acquisition dates.

	Impervious Surface	Non-Forest Vegetation	Forest Canopy	Water
1984–1988	58.09%	−8.08%	−13.01%	−3.13%
1988–1990	−4.21%	23.56%	−6.95%	0.19%
1990–1995	10.05%	4.16%	−7.25%	−1.15%
1995–2000	−15.62%	9.53%	3.95%	2.27%
2000–2004	2.40%	15.95%	−10.90%	0.40%
2004–2009	22.93%	−7.10%	−7.87%	1.41%
2009–2014	17.75%	−14.69%	−2.06%	0.04%
2014–2019	−1.19%	−7.88%	6.94%	−2.83%
2019–2021	13.55%	5.23%	−14.36%	−2.68%

. As indicated by the final land cover class area estimations, since 1984, Chattanooga has steadily lost forest canopy. In the last 37 years between 1984 and 2021, this study estimates that Chattanooga has lost approximately 36.9 sq mi of its urban forest canopy. These losses in forest canopy are replaced by steady gains in impervious surface areas. Since 1984 Chattanooga’s urban spaces have gained 32.5 sq mi, an increase of approximately 134%.

The most significant change in impervious surface areas between consecutive images occurred between 1984 and 1988. During this time, Chattanooga’s urban spaces increased in area by approximately 58%, while urban forest area decreased by 13%. The most significant change in urban forest canopy surface area occurred between 2019 and 2021. During this period, urban forest areas decreased by 14%, while impervious surface area increased by 14%. In two separate instances, the percent change in urban forest canopy is positive rather than negative, suggesting a net growth in urban forest canopy. The first instance, occurring between 1995 and 2000, is complemented by an increase of approximately 10% of non-forest vegetation.

In recent years, the City of Chattanooga has undergone steady population growth, mostly due to rapid economic growth [60], which is heavily favored by its geographic location. The city is home to several large organizations such as Volkswagen, Unum, Tennessee Valley Authority, Blue Cross Blue Shield, Wacker, and Amazon that are driving the area’s economic growth. The economic growth of the city is strongly attributed to the implementation of the fiber optic internet by the City’s Electric Power Board (EPB). With numerous large corporations continuing to expand, a nationally ranked internet infrastructure, and a supported nickname as the “Scenic City”, Chattanooga’s economic and social environment is becoming increasingly attractive to startup businesses. The city has also designated a large part of the downtown area for startups, small businesses, nonprofits, and government offices called the “Innovation District” [61]. These clearly indicate the underlying reasons for the continuous growth of urban areas in Chattanooga, and thereby support the trends of changes in urban forest canopy revealed in this study.

3.3. Accuracy Assessment

For each of the classified land cover datasets, a confusion matrix was generated (

Table 6. Confusion matrices derived from the accuracy assessment of 1984–2000 land cover datasets.

	Classified Pixels	Developed	Non-Forest Vegetation	Forest	Water	Total (Classified)	User’s Accuracy
1984	Developed Pixels	413	18	2	0	433	95.38%
	Non-Forest Vegetation Pixels	166	206	135	0	507	40.63%
	Forest Pixels	0	0	1026	0	1026	100.00%
	Water Pixels	0	0	0	2599	2599	100.00%
	Total (Reference)	579	224	1163	2599	4565
	Producer’s Accuracy	71.33%	91.96%	88.22%	100.00%		Overall = 92.97%
	Classified Pixels	Developed	Non-Forest Vegetation	Forest	Water	Total (Classified)	User’s Accuracy
1988	Developed Pixels	214	3	3	0	220	97.27%
	Non-Forest Vegetation Pixels	66	172	60	0	298	57.72%
	Forest Pixels	22	28	709	0	759	93.41%
	Water Pixels	0	0	0	2599	2599	100.00%
	Total (Reference)	302	203	772	2599	3876
	Producer’s Accuracy	70.86%	84.73%	91.84%	100.00%		Overall = 95.30%
	Classified Pixels	Developed	Non-Forest Vegetation	Forest	Water	Total (Classified)	User’s Accuracy
1990	Developed Pixels	206	0	0	0	206	100.00%
	Non-Forest Vegetation Pixels	102	263	6	0	371	70.89%
	Forest Pixels	0	0	802	0	802	100.00%
	Water Pixels	0	1	1	2597	2599	99.92%
	Total (Reference)	308	264	809	2597	3978
	Producer’s Accuracy	66.88%	99.62%	99.13%	100.00%		Overall = 97.23%
	Classified Pixels	Developed	Non-Forest Vegetation	Forest	Water	Total (Classified)	User’s Accuracy
1995	Developed Pixels	157	0	0	0	157	100.00%
	Non-Forest Vegetation Pixels	2	294	43	0	339	86.73%
	Forest Pixels	0	0	431	0	431	100.00%
	Water Pixels	0	0	0	2599	2599	100.00%
	Total (Reference)	159	294	474	2599	3526
	Producer’s Accuracy	98.74%	100.00%	90.93%	100.00%		Overall = 98.72%
	Classified Pixels	Developed	Non-Forest Vegetation	Forest	Water	Total (Classified)	User’s Accuracy
2000	Developed Pixels	132	0	0	0	132	100.00%
	Non-Forest Vegetation Pixels	0	164	149	0	313	52.40%
	Forest Pixels	41	0	606	0	647	93.66%
	Water Pixels	0	0	0	2599	2599	100.00%
	Total (Reference)	173	164	755	2599	3691
	Producer’s Accuracy	76.30%	100.00%	80.26%	100.00%		Overall = 94.85%

and

Table 7. Confusion matrices derived from the accuracy assessment of 2004–2021 land cover datasets.

	Classified Pixels	Developed	Non-Forest Vegetation	Forest	Water	Total (Classified)	User’s Accuracy
2004	Developed Pixels	125	0	0	0	125	100.00%
	Non-Forest Vegetation pixels	51	187	0	0	238	78.57%
	Forest Pixels	0	1	417	0	418	99.76%
	Water Pixels	0	0	0	2599	2599	100.00%
	Total (Reference)	176	188	417	2599	3380
	Producer’s Accuracy	71.02%	99.47%	100.00%	100.00%		Overall = 98.46%
	Classified Pixels	Developed	Non-Forest Vegetation	Forest	Water	Total (Classified)	User’s Accuracy
2009	Developed Pixels	295	0	0	0	295	100.00%
	Non-Forest Vegetation Pixels	24	159	10	0	193	82.38%
	Forest Pixels	0	0	1658	0	1658	100.00%
	Water Pixels	0	0	0	2599	2599	100.00%
	Total (Reference)	319	159	1668	2599	4745
	Producer’s Accuracy	92.48%	100.00%	99.40%	100.00%		Overall = 99.28%
	Classified Pixels	Developed	Non-Forest Vegetation	Forest	Water	Total (Classified)	User’s Accuracy
2014	Developed Pixels	297	0	0	0	297	100.00%
	Non-Forest Vegetation Pixels	46	161	5	0	212	75.94%
	Forest Pixels	1	0	1190	0	1191	99.92%
	Water Pixels	0	0	0	2599	2599	100.00%
	Total (Reference)	344	161	1195	2599	4299
	Producer’s Accuracy	86.34%	100.00%	99.58%	100.00%		Overall = 98.79%
	Classified Pixels	Developed	Non-Forest Vegetation	Forest	Water	Total (Classified)	User’s Accuracy
2019	Developed Pixels	253	0	0	0	253	100.00%
	Non-Forest Vegetation Pixels	15	152	6	0	173	87.86%
	Forest Pixels	0	8	1288	0	1296	99.38%
	Water Pixels	0	0	0	2599	2599	100.00%
	Total (Reference)	268	160	1294	2599	4321
	Producer’s Accuracy	94.40%	95.00%	99.54%	100.00%		Overall = 99.33%
	Classified Pixels	Developed	Non-Forest Vegetation	Forest	Water	Total (Classified)	User’s Accuracy
2021	Developed Pixels	586	1	0	0	587	99.83%
	Non-Forest Vegetation Pixels	3	302	0	0	305	99.02%
	Forest Pixels	0	13	2431	0	2444	99.47%
	Water Pixels	0	0	0	2599	2599	100.00%
	Total (Reference)	589	316	2431	2599	5935
	Producer’s Accuracy	99.49%	95.57%	100.00%	100.00%		Overall = 99.71%

). Based on the results of the accuracy assessments conducted for the ten land cover datasets, overall accuracy ranged from 92.97% to 99.71%. Scene 10 (2021) yielded the highest overall accuracy, while Scene 1 (1984) produced the lowest. Results across all ten accuracy assessments suggest the most common error of commission (Type I error/false-positive) was the erroneous classification of non-forest vegetation pixels into the impervious surface and forest canopy land classes. The most common error of omission (Type II error/false-negative) was the erroneous classification of impervious surface pixels into the non-forest vegetation and forest canopy land classes. Based on the results from the accuracy assessments, the methodology utilized in this research is capable of classifying forest canopy pixels apart from other land cover classes using moderate resolution Landsat 5 TM imagery and Landsat 8 OLI imagery at the city-scale with substantial accuracy.

4. Discussion

The main objective of this research was to identify the extent of urban forest across the City of Chattanooga, TN. In conducting this research, Landsat 5 and 8 data were obtained. However, Landsat 5 and 8 are two different systems designed three decades apart. Therefore, image quality between these systems is inherently different. Additionally, variable levels of atmospheric dust, pollen, and water vapor can impact the accuracy of a remote sensor. For this reason, a supervised hybrid classification workflow as described in this research was implemented in lieu of more traditional classification methods to normalize potential variation across classified results. In addition, the use of Level-2 calibrated surface reflectance images should minimize the uncertainty between the two sensor systems. However, it is recommended to apply more advanced machine learning techniques such as Artificial Neural Network (ANN) and/or Deep Learning (DL) models to expand this research further in the future.

To accurately classify the extent of forest canopy in each of the images, it is necessary to delineate tree pixels from pixels representing non-tree types of vegetation. However, because of Landsat’s moderate 30 m spatial resolution, and the spectral similarities between tree pixels and non-tree pixels, it is highly difficult to visually differentiate between forest canopy and non-forest vegetation, even when viewing the imagery in false color. This was especially true in residential areas. Similarly, many developed residential areas in Chattanooga have higher quantities of forest canopy compared to developed non-residential areas in the city. However, the forest canopy in these residential areas is unlike a natural forested area. In residential areas, the forest canopy is highly fragmented and mixed in thoroughly with other non-forest vegetation and impervious surfaces. As a result, it becomes hard to draw the line between forest canopy, non-forest vegetation, and developed land areas within residential areas throughout Chattanooga. In other words, due to Landsat’s limited 30 m spatial resolution and the spectral similarities between tree pixels and non-tree vegetation pixels, it is difficult to visually differentiate between forest and non-forest vegetation, especially in residential areas, where fragments of forest, non-forest vegetation, and developed land classes are packed together in an area smaller than Landsat’s spatial resolution. Therefore, this research opted to utilize unsupervised classification for pervious pixel classification, as there was a greater chance for the computer to spectrally differentiate between forest and non-forest vegetation pixels. The suggested use of ANN and/or DL algorithms potentially should resolve this issue further.

Ultimately, the issues as described here can be tied back to the spatial resolution of Landsat, but also to the scale of the objects being classified. Because many of the non-forest vegetation pixels represent vegetation that has been re-planted by people physically within developed areas, and these re-planted areas of non-forest vegetation are typically smaller than 30 m, such as, for example, a flower garden or a grassy lawn, it is possible for non-forest vegetation pixels to blend with the adjacent developed and forest canopy pixels due to the moderate spatial resolution of Landsat. The pan-sharpening technique can be useful in this regard for Landsat 7 and Landsat 8. However, since Landsat 5 does not have a panchromatic band, this approach cannot be used for the entire time series.

Object-based supervised classification is a relatively new concept that groups spectrally similar and spatially clustered pixels into objects or segments, then conducts classification on the segments. One caveat to object-based classification is that it is typically utilized with fine spatial resolution sensors. Additionally, segmentation is a computationally demanding process that requires a distributed processing environment. However, it is strongly recommended to apply object-based supervised classification to expand this research further to see if this approach could generate better results.

Cloud coverage is always a challenge for any optical sensors. Although this research only used Landsat imagery, future research should consider using imagery from other similar optical sensors at least for the recent dates of the time series to minimize the possibility of large data gaps.

To future researchers conducting spatial analyses, when building temporal datasets for spatiotemporal analyses, it is often common to attempt to acquire as much data as possible within a defined acquisition period. However, when working with many multispectral datasets, generating classified outputs that are similar enough to each other that, when viewed chronologically, the under classification or over classification of a given class is not directly obvious is quite difficult. Therefore, it could be helpful to only acquire data around the beginning and end of the defined temporal scope of a research project, as the net loss/gained, and percent change is commonly the information desired by the researcher and other stakeholders.

The main objective of this research was to detect, map, and quantify the historical coverage of the urban forest in Chattanooga, TN. To reduce confusion, this research did not expand the analysis further. However, it is recommended to add more analysis to expand this research further in the future. In addition, it would also be useful to include analysis of how different types of land use and land cover during the urbanization process specifically affect the forest ecosystem.

5. Conclusions

This research was designed to detect and map the historic extent of Chattanooga’s urban forest canopy from 1984 to 2021 using remote sensing technology. This research took advantage of the extensive temporal archive of multispectral satellite imagery provided by the Landsat program to conduct a 37-year land cover change analysis across Chattanooga, Tennessee.

This study shows that since 1984 the forest canopy in Chattanooga regularly decreased. During the 37-year time span between 1984 and 2021, Chattanooga lost approximately 36.9 sq mi of its urban forest canopy. These losses in the forest canopy were replaced by steady gains in impervious surface areas. Since 1984, Chattanooga’s urban spaces have gained 32.5 sq mi, an increase of approximately 134%.

Based on the obtained results, it can be concluded that, despite the limitations in spatial resolution, Landsat satellite images can be effectively used for mapping and analyzing the multidecade spatiotemporal forest canopy in a rapidly expanding metropolitan center, such as Chattanooga, TN. The data generated in this research and the developed image processing method have great potential to help direct sustainable development efforts towards areas urbanizing at an above-average rate.