remote sensing

Article
Classification of Urban Area Using Multispectral
Indices for Urban Planning
Philip Lynch 1, * , Leonhard Blesius 1 and Ellen Hines 1,2
1 Department of Geography and Environment, San Francisco State University, 1600 Holloway Ave.,
San Francisco, CA 94132, USA; lblesius@sfsu.edu (L.B.); ehines@sfsu.edu (E.H.)
2 Estuary and Ocean Science Center, San Francisco State University, 3150 Paradise Dr., Tiburon, CA 94920, USA
* Correspondence: plynch15@yorku.ca; Tel.: +1-415-338-2049




Received: 14 July 2020; Accepted: 2 August 2020; Published: 4 August 2020


Abstract: An accelerating trend of global urbanization accompanying population growth makes

frequently updated land use and land cover (LULC) maps critical. LULC maps have been widely
created through the classification of remotely sensed imagery. Maps of urban areas have been both
dichotomous (urban or non-urban) and entailing of discrete urban types. This study incorporated
multispectral built-up indices, designed to enhance satellite imagery, for introducing new urban
classification schemes. The indices examined are the new built-up index (NBI), the built-up area
extraction index (BAEI), and the normalized difference concrete condition index (NDCCI). Landsat
Level-2 data covering the city of Miami, FL, USA was leveraged with geographic data from the Florida
Geospatial Data Library and Florida Department of Environmental Protection to develop and validate
new methods of supervised and unsupervised classification of urban area. NBI was used to extract
discrete urban features through object-oriented image analysis. BAEI was found to possess properties
for visualizing and tracking urban development as a low-high gradient. NDCCI was composited
with NBI and BAEI as the basis for a robust urban intensity classification scheme superior to that of
the United States Geological Survey National Land Cover Database 2016. BAEI, implemented as a
shadow index, was incorporated in a novel infill geosimulation of high-rise construction. The findings
suggest that the proposed classification schemes are advantageous to the process of creating more
detailed cartography in response to the increasing global demand.

Keywords: cartography; global urbanization; urban remote sensing; geographic information systems;
geosimulation; land change science; geospatial analysis; infill development

1. Introduction
Today, the world’s rapidly developing trend of urbanization has made frequently updated
surface maps critical [1–4]. Land use and land cover (LULC) maps are advantageous for purposes
of government planning, environmental management, disaster management, and education of the
general public on the status of global development. The impacts of urbanization for the environment
include triggering potentially harmful feedback such as climate change, reducing water quality,
and the replacement of nature by human construction [5–7]. The link between urbanization and
environmental impacts can be analyzed by mapping their extent and severity in relation to urban
expansion [8–10]. LULC maps may serve as tools of emergency response to disasters such as fires,
earthquakes, and floods where the extent and severity of the disasters can be displayed and analyzed
to support response measures. The problem of producing these maps on a large scale can be solved
with satellite remote sensing.
Ordinarily, satellite remote sensors feature multiple spectral bands for use in analysis, where each
band may be advantageous given the properties of materials identifiable in different parts of the light

Remote Sens. 2020, 12, 2503; doi:10.3390/rs12152503 www.mdpi.com/journal/remotesensing

Remote Sens. 2020, 12, 2503 2 of 28

spectrum [11–21]. Satellite remote sensors take a bird’s-eye view of changes related to urban growth.
While high-resolution (1 × 1 m) remote sensing has proven to be very useful for mapping the extent
of urbanization as well as for mapping the presence of individual feature footprints, the research
presented here focuses primarily on satellites of moderate resolution (10 × 10–30 × 30 m) [12,21,22].
Moderate-resolution data may be sufficient for mapping urban extent and is typically available for free
in large time-series volumes compared to high-resolution data which is expensive to acquire, making it
unsuitable for institutions with limited budgets.
Typically, satellite image analysts take a dichotomous approach to LULC mapping, classifying
image cells as either urban or non-urban [11–19]. These maps often depict the encroachment of urban
sprawl expanding into the natural landscape. Still, some cities do not grow outwardly because of
physical constraints (other cities, terrain, water bodies, etc.) or should not grow outwardly because
of planning regulations. The world population is expected to increase from ~8 billion currently to at
least 10 billion in three decades, driving the impact of problems associated with urbanization to new
levels [4,23]. In response, governments are enacting smart growth policies meant to limit outwardly
expanding cityscapes into compact, walkable urban areas with higher populations [3]. Smart growth
implies that environmental impacts are lessened by condensing urban infrastructure [3]. Therefore,
the drawing of infill development policies is the foremost solution to the problems associated with
the outward expansion of cityscapes against the natural landscape [3]. The term infill development
refers to the rededication of urban land to new construction. Simply, urban growth takes place within
the confines of the cityscape, such as the development of vacant lots, redevelopment of run-down
neighborhoods, and conversion of parks to construction.
An inventory of urban land can be obtained through field methods as well as the classification of
aerial photographs and satellite imagery and cataloged as data in a geographic information system
(GIS). Classifying every land use plot specifically as possible may take a long time, where the necessary
geographic view can be achieved more efficiently through a large-scale remote sensing classification.
Numerous attempts to classify urban land use by discrete types are on record in scientific journals.
Among the methods that may be implemented to create such maps are iterative unsupervised
classification, such as k-means and iso cluster, supervised classification requiring class training sets,
such as parametric Gaussian maximum likelihood (GML) or non-parametric support vector machine
(SVM), and object-oriented image analysis (OBIA) [24–28]. While classifiers by themselves are used
to create maps featuring multiple urban types, the process can be refined and simplified using
multispectral unsupervised classification formulas called spectral enhancements.
Remote sensing spectral enhancements are formulas used to transform multispectral data retrieved
from a sensor into indexed quantities useful for analyzing a target environment, isolating target features,
and indexing those features into identifiable classes [8–20,27,29]. For instance, measurements of solar
radiation reflecting from Earth’s surface to a satellite sensor can be manipulated to yield a description
of plant biology, the presence of bare soil, or a state of urbanization [12]. For example, Zha et al.
(2003) [18] introduced a method for extracting urban features from satellite data using a spectral
enhancement. The normalized difference built-up index was used to extract urban features of Nanjing,
China from Landsat 5 Thematic Mapper (TM) 30 × 30 m resolution data. It enhances the spectral
response differences that exist between shortwave infrared (SWIR) and near-infrared (NIR) bands
to separate urban from non-urban areas in an image. The authors reported an accuracy of 92.6%
based on an assessment of 68 randomly distributed points in their study area and assert that spectral
enhancements increase urban feature classification accuracy compared to the GML classifier.
Ettehadi et al. (2019) [27] mapped the land cover of Istanbul by classifying enhanced images,
derived from Sentinel-2A 10 × 10–20 × 20 m data, using SVM. The normalized difference tillage
index was used to highlight urban features as part of a three-band composite image, including the
red-edge-based normalized difference vegetation index, and the modified normalized difference water
index. They delineate asphalt, industrial land use, and other built-up as part of their classification
scheme. The reported producer and user accuracies are 85.71% and 100% for asphalt areas, 98.28%
Remote Sens. 2020, 12, 2503 3 of 28

and 81.43% for industrial areas, and 93.28% and 92.25% for other built-up. It is of critical notice
that differences in environmental factors like lighting, the complexity of land cover, etc. that occur
between regions will impact the performance of enhancements [12]. While there is record of successful
methodology for classifying discrete urban land use types with moderate resolution data, in addition
to methodology for extracting feature footprints from high-resolution satellite imagery, research
still needs to be conducted for extracting the footprints of urban land use geometries (recreational,
residential, business, etc.) from moderate resolution data [21,22,27]. Additionally, though the state
of environmental variables are frequently measured with low-high gradients, such as the use of the
well-known normalized difference vegetation index (NDVI) to measure plant vigor, there is no volume
of knowledge available about measuring the state of urban development by gradient [8–10,29].
The United States Geological Survey (USGS) publishes the National Land Cover Database (NLCD),
derived from Landsat data and delineating all land in the conterminous United States into 21 classes,
every five years. NLCD features four classes of urban intensity derived entirely from the NLCD
Percent Developed Imperviousness product, also released on the same five-year basis. The classes
are open space, low intensity, medium intensity, and high intensity. Open space is typified by mostly
vegetation (<20% impervious), low intensity by single-family homes (20–49% impervious), medium
intensity by single-family homes (50–79%), and high intensity by frequently traversed areas (80–100%
impervious) [30]. These urban intensities are based on groupings of developed surface imperviousness
derived from satellite data, which physically limits the interpretation of feature structure. There is
room in scientific writing for a methodology, emphasizing automation, for correcting the linear errors
associated with deriving classes from only a satellite assessment of surface imperviousness.
In addition to static classification maps, urban growth models, called geosimulations, are used
to generate land cover maps of predictable conditions by assigning transition probabilities and
potentials to maps derived from remote sensing [1–3]. Rahimi (2016) [3] utilized a combination of GIS,
remote sensing, and a multilayer perceptron (MLP) artificial neural network (ANN) in an attempt
to map potential infill development for the city of Tabriz, Iran. Urban land cover was classified,
cell-by-cell, as either urban or subject to infill development. LULC classifications were derived from
Landsat 5 data for 1989 and Satellite Pour l’Observation de la Terre (SPOT) 10 × 10 m data for 2005.
The classifications were cross-referenced with geographic predictor variables such as distance from
roads and population density using an MLP ANN geosimulation, thus predicting urban land cover for
2021. Sun et al. (2007) [1] modeled the growth of Calgary, Canada, using a cellular automata/Markov
chain combination model. They predicted change, using classified Landsat 5 data from 1985 and 1992,
regarding the presence of residential and commercial, industrial, transportation, and park development
compared to vacant areas and water bodies. Suitability maps of Calgary’s “Future Conceptual Urban
Structure” were created for each class, merged, and used as a predictor. The researchers were able to
achieve favorable results according to a cellular accuracy assessment using data of known conditions
of the predicted year, 1999. Where geosimulations of varying accuracy are on record, a knowledge gap
exists concerning the difficulty of predicting discrete changes, such as the construction of skyscrapers,
on a cell-by-cell basis [1–3]. There should be research of methodology for geosimulating a prediction
encompassing the formation of such changes.
The purpose of this paper is to advance our current understanding of the processes of leveraging
remote sensing to identify discrete urban land use types for incorporation into modeling schemes
such as static mapping and geosimulations. The study evaluated the possibility of implementing
spectral enhancements designed to enhance construction features from the natural landscape for the
creation of new classification schemes that delineate urban land use types. It seeks to establish means
of separating urban features captured in enhancements by distinguishing identifiable attributes such
as albedo, the scale of development, and surface configuration. Today’s GIS software packages include
intuitive tools for performing the necessary geospatial analyses such as pixel-based image classification
and OBIA, time-series change detection, and conditional geoprocessing. The aim is to propose new
schemes, based on enhancements, for classifying urban land using:
Remote Sens. 2020, 12, 2503 4 of 28

1. Classification by object: Urban land use types may be identifiable by differences in albedo
attributed to the presence of varying construction materials, vegetation, and bare ground.
2. Classification by gradient: The amount of radiation absorbed by urban surfaces may fluctuate
due to deconstruction (i.e., by natural disasters) or the accumulation of construction material
(i.e., by infill development).
3. Classification by intensity: Urban feature structure may be mapped by compositing the properties
of albedo, absorbed radiation, and percent of developed imperviousness.
4. Classification by focalized transition potential; development trends attributed to discrete
occurences identifiable in imagery may be generalized to accurately predict the extent of
those occurences.

The enhancements analyzed are the new built-up index (NBI) by Jieli et al. (2010) [14], the built-up
area extraction index (BAEI) by Bouzekri et al. (2015) [12], and the normalized difference concrete
condition index (NDCCI) by Samsudin et al. (2016) [20]. NBI yields a low-high assessment of surface
albedo, which may be useful for discriminating groups of urban features according to their relative
brightness. BAEI works as a view of the surface net radiation balance (NRB), wherefore it may be
possible to measure the amount of construction material present on the surface. The NDCCI band
configuration differs slightly from NDVI, designed to focus on construction instead of vegetation,
which may make it more useful for urban mapping. Urban intensities published by the USGS in NLCD
2016 are analyzed for classification errors, and a corrected classification was modeled with the NLCD
2016 Percent Developed Imperviousness product in tandem with SVM supervised and iso cluster
unsupervised classifications of an NBI/BAEI/NDCCI composite. A novel geosimulation of high-rise
development was performed with land use classifications derived from the BAEI as a shadow index,
based on focalized index values integrated with manual classifications and using polynomial transition
trends as predictors.

2. Methods
ESRI’s ArcGIS [31] was the software used in most of the study; for geoprocessing, geospatial
analysis, and cartography. ERDAS Imagine [32] was used for image histogram analysis to establish
BAEI feature extraction thresholds. Clark Lab’s TerrSet [33] was used for geosimulating future land
use with the land change modeler (LCM) module for MLP ANN Markov chain analysis. The research
was conducted on a place with satellite coverage, where urban growth modeling is highly relevant to
environmental sustainability and can be validated using a sufficient volume of ground truth data.

2.1. Study Area

Miami, FL, USA is located on the southeastern tip of Florida at 25◦ 460 29.8128”N/80◦ 110 9.5604”W,
spans ~143 sq km, and serves as the Miami-Dade County seat (see Figure 1). The land use types
comprising the city are primarily residential, commercial, and industrial. Miami poses a unique
challenge to the science of urban remote sensing. It does not grow outwardly because other cities and
development border its western periphery, and its eastern boundary is the coastline. It is a rapidly
developing metropolis with a recorded population increase of ~350,000 in 1980 to ~486,000 in 2018 [34].
Miami-Dade County has enacted development concurrency policies meant to force a compact urban
design by slowing sprawl and promoting infill development [35]. For instance, some areas of Miami
became hotspots for construction of multifamily housing between 1995–2004, after the inception of
Transportation Concurrency Exception Areas that allow for infill development despite traffic and
accessibility limits [35]. A coastal setting containing sensitive marine environments makes this city a
highly relevant target for an inquisition into environmental conservancy by actively managing urban
growth. If remote sensing is implemented as a means of monitoring Miami’s growth, it should be
modeled to account for upward growth. A collection of cloud-free Landsat data is available from the
USGS in addition to volumes of GIS data provided by the state of Florida and Miami-Dade County.
 Remote Sens. 2020, 12, x FOR PEER REVIEW 5 of 29

Remote Sens. 2020,

available from12,the
2503USGS in addition to volumes of GIS data provided by the state of Florida 5and
of 28
Miami-Dade County.

Figure1.1.Miami
Figure Miamisits
sitson
onthe
thesoutheastern
southeastern tip
tip of
of Florida.
Florida. The
The city
city is
is enclosed
enclosed by
by Biscayne
BiscayneBay
Bayand
andthe
the
Atlantic Ocean to the east, and urbanized parts of Miami-Dade County
Atlantic Ocean to the east, and urbanized parts of Miami-Dade County inland. inland.

2.2.
2.2.Data
DataSources
Sources
Vector
VectorGISGISdata
dataofofMiami
Miamiwerewere retrieved
retrieved from
from Miami-Dade
Miami-Dade County Open GIS.
County Open GIS. That
That dataset
dataset
includes
includesshapes
shapesof ofthe
theMiami
Miamimunicipal
municipal boundary,
boundary, county
county water
water features, and the the shoreline.
shoreline. Florida
Florida
neighborhood
neighborhoodshapesshapes werewere retrieved
retrieved from from Zillow
Zillow Real
Real Estate
Estate Listings. Finally, shapes
Listings. Finally, shapes of of Florida
Florida
LULC
LULCdesignations
designationswere wereretrieved
retrievedfromfromthe the Florida
Florida Geographic
Geographic Data Library (FGDL) (FGDL) andandthe
theFlorida
Florida
Department
DepartmentofofEnvironmental
EnvironmentalProtection
Protection(FDEP)(FDEP) Geospatial
Geospatial Open Data. The The LULC
LULCdesignation
designationfilesfiles
wereconverted
were convertedtoto30 30××3030mmraster
raster format
format to to facilitate
facilitate cell-by-cell analyses.
analyses. TheTheGIS
GISdata
dataareareused
usedto to
clipand
clip andanalyze
analyzeselect
selectportions
portions of of satellite
satellite data
data acquired
acquired from the National
National Aeronautics
Aeronauticsand andSpace
Space
Administration(NASA)
Administration (NASA)Landsat
Landsatsatellite
satelliteseries.
series. NLCD
NLCD 2016 data were published by by the
the USGS.
USGS.
Accordingtotothe
According theUSGS
USGS[36],
[36],the
the3030××30 30mmLandsat
LandsatSurface
SurfaceReflectance
ReflectanceLevel-2
Level-2science
sciencedata
datathat
thatis
is available free to the public from their EarthExplorer website is designed to support
available free to the public from their EarthExplorer website is designed to support the analyses of land the analyses of
land change
change science.science.
Level-2 dataLevel-2 data is preprocessed
is preprocessed with both georeferencing
with both georeferencing and atmospheric and corrections
atmospheric to
corrections
normalize to normalize
every image in the every image
dataset for in the dataset
comparison forLevel-2
[36]. comparison
Landsat[36]. Level-2
5 data Landsat
receive 5 data
atmospheric
receive atmospheric
corrections in the form corrections in theSimulation
of 6S (Second form of 6Sof (Second Simulation
a Satellite Signal in ofthe
a Satellite Signal in the
Solar Spectrum) Solar
meant to
Spectrum)
minimize themeant to minimize
influence of waterthe influence
vapor, ozone, ofgeopotential
water vapor,height,
ozone, aerosol,
geopotential height, aerosol,
and elevation and
on spectral
elevation
returns [36].onSimilarly,
spectral Landsat
returns [36]. Similarly,Land
8 Operational Landsat 8 Operational
Imager (OLI) data is Land Imagerwith
corrected (OLI) data is
an internal
corrected with an internal satellite algorithm [36]. All Level-2 images used
satellite algorithm [36]. All Level-2 images used for this study are georeferenced to the Universal for this study are
georeferenced
Transverse to the(UTM)
Mercator Universal
ZoneTransverse
17N projected Mercator (UTM)system
coordinate Zone 17N[36].projected coordinate
Subsequently, system
all other data
[36]. Subsequently, all other data was initially projected to this system to
was initially projected to this system to facilitate analysis, except for the NLCD products. facilitate analysis, except for
the The
NLCD products.
imagery used in this study includes Landsat 5 data for the dates of 2 November 1985,
6 November 1998, 24 October 1999, 20 November 2003, 17 November 2008, 19 October 2009,
2 November 2011, and Landsat 8 data for the dates of 17 October 2014 and 22 October 2016. The temporal
Remote Sens. 2020, 12, 2503 6 of 28

sequencing of the data was intended to minimize the spectral influence of differences that will occur
because of seasonal conditions. Due to sensor saturation, the data may contain minor errors that pitch
some cell values outside of the 0-1 range of proper reflectance values. These cells are screened out of
each band during geoprocessing to preserve analytical accuracy. Table 1 lists all of the GIS and remote
sensing data used, including the year or date covered by the data, sources, file formats, and the amount
of land cloud cover in imagery as estimated by the USGS.

Table 1. Inventory of all of the geospatial data used during this study, including descriptive information.

GIS Data
Description Year(s) Source Format
Miami-Dade Municipal Boundaries
Miami-Dade Municipal Coastal Boundary Miami-Dade Open
2018
GIS
Miami-Dade Water Features Vector shapefiles
Zillow Real Estate
Florida Neighborhood Boundaries 2018
Listings
Generalized Land Use Derived from 2017
Florida Geographic
Parcels - Florida Department of Transportation 2017
Data Library
District 6
Vector shapefiles
Florida Department (converted to raster
of Environmental images)
Statewide Land Use Land Cover 2014–2016
Protection Geospatial
Open Data
National Land Cover Database
2016 United States Raster images
National Land Cover Database Percent
Geological Survey
Developed Imperviousness
Remote Sensing Data
Landsat Scene ID Date Cloud Cover Sensor
LT05_L1TP_015042_19851102_20161004_01_T1 2 November 1985 0%
LT05_L1TP_015042_19981106_20160922_01_T1 6 November 1998 0%
LT05_L1TP_015042_19991024_20160919_01_T1 24 October 1999 0%
LT05_L1TP_015042_20031120_20160914_01_T1 20 November 2003 0% Landsat 5 TM
LT05_L1TP_015042_20081117_20160908_01_T1 17 November 2008 0%
LT05_L1TP_015042_20091019_20160904_01_T1 19 October 2009 2%
LT05_L1TP_015042_20111110_20160830_01_T1 2 November 2011 0%
LC08_L1TP_015042_20141017_20170303_01_T1 17 October 2014 0.01%
Landsat 8 OLI
LC08_L1TP_015042_20161022_20170219_01_T1 22 October 16 0.08%

2.3. Evaluation of Published Built-Up Indices

Landsat data were used to evaluate eleven enhancements by visual assessment of the study area,
including the adjustment of symbology and display settings. The enhancements were selected through
rigorous review of scientific literature relevant to urban feature extraction, where each index was found
to be well-known or recently published. Nine were designed to extract urban features from moderate
resolution data, one designed to extract urban features from SPOT data, and one designed to assess the
quality of building materials with high-resolution data. The enhancements were initially derived from
the 2016 Landsat 8 image, and a time-series incorporating all of the images was evaluated for BAEI.
A list of these enhancements is given in Table 2. The three enhancements selected for further analysis,
based on the identification of unique spectral properties previously undocumented in published
research, were NBI, BAEI, and NDCCI (see Figure 2).
 Remote Sens. 2020, 12, x FOR PEER REVIEW 7 of 29

Table 2. Built-up indices found in scientific journals evaluated for unique spectral properties.
Remote Sens. 2020, 12, 2503 7 of 28
Name of Index Author(s) Formula Accuracy Study Area
The Enhanced
and 2. Built-up
Built-up Table As-syakur indices
et found in scientific journals evaluated for unique spectral properties.
(SWIR1-NIR)/10*SQRT Densapar,
88.98%
Bareness Index al. (2012) [11] (SWIR1+THERMAL) Indonesia
Name of Index Author(s) Formula Accuracy Study Area
(EBBI)
The Enhanced
The Built-up
Built-upand As-syakur
Bouzekri et et al. (SWIR1−NIR)/10*SQRT
88.98%
Densapar,
AreaBareness
Extraction Index (2012) [11] (RED+.3)/(GREEN+SWIR1)
(SWIR1+THERMAL) 92.66% Djelfa, Algeria
Indonesia
al. (2015) [12]
Index (BAEI)
(EBBI)
The
TheBuilt-up
Built-up Area National
HeBouzekri
et al. et al. (SWIR1-NIR)/(SWIR1+NIR)-
Extraction
Index Index (RED+0.3)/(GREEN+SWIR1) 92.66%
86.3% Djelfa, Algeria
Olympic Park,
(BAEI) (2010)(2015)
[13] [12] (NIR-RED)/(NIR+RED)
(BUI) Beijing, China
TheTheNew Built-Index National
Built-up Heetetal.
Jieli al. (2010) (SWIR1−NIR)/(SWIR1+NIR)− Changzhou
86.3% Olympic Park,City,
up Index (BUI) [13] (RED-SWIR1)/(NIR)
(NIR−RED)/(NIR+RED) 90%
(2010) [14] Beijing,China
China
(NBI)
The New Built-up
The Dry Built-up RasulJieliet
etal.
al. (2010)(BLUE-THERMAL)/(BLUE+THERMAL)- Changzhou City,
Index (RED−SWIR1)/(NIR) 90%
93% Erbil, Iraq
Index (DBI) (2018) [14]
[15] (NIR-RED)/(NIR+RED) China
(NBI)
The The
Normalized
Dry Built-up Rasul et al. (2018) (BLUE−THERMAL)/(BLUE+THERMAL)−
[(SWIR2-SWIR1)/GREEN]/ 93% Erbil, Iraq
Built-upIndexArea
(DBI) [15] (NIR−RED)/(NIR+RED) 86.87%
[(SWIR2+SWIR1)/GREEN]
Index
The(NBAI)
Normalized Waqar et al. Islamabad,
The Band Ratio (2012) [16] [(SWIR2−SWIR1)/GREEN]/ Pakistan
Built-up Area 86.87%
[(SWIR2+SWIR1)/GREEN]
Index (NBAI)
for Built-up Area Waqar et al. (RED)/(SWIR1) 85.09% Islamabad,
(BRBA)
The Band Ratio for (2012) [16] Pakistan
Built-up Area (RED)/(SWIR1)
2*SWIR1/(SWIR1+NIR)-[NIR/(NIR-RED)+ 85.09%
(BRBA)
The Index-based GREEN/(GREEN+SWIR1)]/
Xu (2008) 2*SWIR1/(SWIR1+NIR)−[NIR/(NIR−RED)+ Fuzhou City,
Built-up Index 2*SWIR1/(SWIR1+NIR)+ 96.77%
[17] GREEN/(GREEN+SWIR1)]/ South China
The(IBI)
Index-based [NIR/(NIR-RED)+GREEN/ Fuzhou City,
Xu (2008) [17] 2*SWIR1/(SWIR1+NIR)+ 96.77%
Built-up Index (IBI) (GREEN+SWIR1)] South China
[NIR/(NIR−RED)+GREEN/
The Normalized (GREEN+SWIR1)]
Zha et al. Nanjing City,
Difference Built-
The Normalized (SWIR1-NIR)/(SWIR1+NIR) 92.6%
(2003)
Zha [18]
et al. (2003) East City,
Nanjing China
up Index (NDBI)
Difference Built-up (SWIR1−NIR)/(SWIR1+NIR) 92.6%
[18] East China
Index (NDBI) Index Designed for Extracting Urban Features from SPOT Data
The Modified BenkouiderIndex Designed for Extracting Urban Features from SPOT Data
95% & Laghouat and
Built-up Area
The Modified et al. (2019) [NIR+(1.57*GREEN)+(2.4*SWIR1)]/(1+NIR)
Benkouider et al. 91%&
95% M’Sila, Algeria
Laghouat and
Index (MBAI)
Built-up Area [19]
(2019) [19]
[NIR+(1.57*GREEN)+(2.4*SWIR1)]/(1+NIR)
91% M’Sila, Algeria
Index (MBAI)
Index Designed for Determining Material Condition from High-resolution Data
The Normalized Index Designed for Determining Material Condition from High-resolution Data
University Putra
Difference
The Normalized University Putra
Samsudin et Malaysia, Seri
Difference
ConcreteConcrete Samsudin et al. (NIR-GREEN)/(NIR+GREEN) 84.44% Malaysia, Seri
(NIR−GREEN)/(NIR+GREEN) 84.44%
Condition
Condition Index al. (2016)
Index
[20][20]
(2016) Kembangan,
Kembangan,
(NDCCI) Malaysia
Malaysia
(NDCCI)

Figure 2. Select spectral enhancements derived from Landsat 8 OLI 22 October 2016 data; (a) NBI,
(b) BAEI, (c) NDCCI.
Remote Sens. 2020, 12, 2503 8 of 28

2.4. Evaluation of the New Built-Up Index for Object-Oriented Urban Feature Extraction
Upon visual assessment, NBI showed pronounced differences between areas of Miami that are
known to possess business infrastructure compared to those that are known to possess no infrastructure
or residential land use. Therefore, it was selected for further evaluation regarding its usefulness for
serving as the basis of an object-oriented urban land use classification scheme. Jieli et al. (2010) [14]
developed this index with Landsat 5 data to automate the process of mapping residential areas of
Changzhou City, China. The researchers report an overall accuracy of 90% based on a survey of 50
random points in their study area.
An index raster derived from the 2016 Landsat 8 data was clipped to the shape of Miami without
feature extraction thresholding. While visual assessment yielded the conclusion that it is easy to spot
differences between one type of urban feature and another in an enhanced image, the point requires
further elaboration. Noticeable differences exist among land cover types captured in satellite data
between the red, NIR, and SWIR1 bands, and the NBI band arrangement yields an unsupervised
classification where the brightest features (typically urban or barren land) have the highest values [14].
Thus, index values for areas that are only urban will show enhanced brightness for features with high
albedos, such as white concrete or the bare ground of construction sites. Subsequently, zonal statistics
describing the central tendency of index values were calculated for each land use type featured in the
FGDL data. It must be noted that FGDL data does not feature polygon geometries for streets and is
advantageous because it contains additional details regarding the presence of industrial development
compared to the data from FDEP. ArcGIS includes the Mean Segment Shift interactive OBIA tool for
grouping parts of an image with similar properties into segments.
Phiri & Morgenroth (2017) [37] define OBIA as the automatic digitization of homogenous image
features. It works by grouping pixels into vectored segments and then assigning a class to like
segments [37]. Segments are defined by spectral, spatial, and geometric properties [37]. Spectral detail
refers to color characteristics, such as the difference between one type of commercial construction
and another. Spatial detail refers to identifiable differences in spatial characteristics, such as street
lines compared to residential blocks. Geometric detail refers to the minimum size of an image
segment, which can aggregate details on, for example, a pixel, block, neighborhood, or city level.
The method is useful for avoiding a scattered, “salt and pepper” look typically associated with
pixel-based classification methods.
The 2016 NBI was segmented into object primitives useful for identifying enhanced color intensity,
object formation, and surface texture. An object encompassing industrial land use in the Little Haiti
neighborhood of Miami was extracted by vectorizing a segment. The object was cross-referenced
with the FGDL data to validate the usefulness of non-automatic, interactive segment selection as a
classification method. While discrete urban types can be classified through OBIA feature extraction
with this index, there is still an additional need for analysis of the scale of development within classes.

2.5. Evaluation of the Built-Up Area Extraction Index for Change Detection of Gradient Development
BAEI provides an enhanced view of the accumulation of construction material as a low-high
gradient. The index was selected for further evaluation regarding its usefulness for time-series change
detection of a quantifiable increase in urban development (different grades of building presence) because
of its range of non-linear values, which may prove useful for tracking shifts in urban morphology.
In addition to a pronounced gradient, the index feature extraction threshold separates high-rise
shadows from the rest of the urban area.
BAEI is a high-precision, non-linear (the positive range of values extends beyond 0 to 1) feature
extraction index developed with Landsat 8 data by Bouzekri et al. (2015) [12] to automate the process
of extracting built-up areas of Djelfa, Algeria. They report an accuracy of 92.66% based on a survey of
50 random points in their study area. Compared to the other built-up indices evaluated, it provides
an expanded range of values that may be useful for classifying urban states with low-high values.
Remote Sens. 2020, 12, 2503 9 of 28

An observable increase in the central tendency of index values for an urban area, over a time-series,
requires additional explanation.
Chrysoulakis (2003) [38] utilizes NASA’s Advanced Spaceborne Thermal Emission and Reflection
Radiometer 15 × 15–30 × 30 m satellite data, in combination with in situ data, to develop an estimation
of the all-wave surface NRB for Athens, Greece. NRB is defined in the formula:

NRB = (1 − αshort)E + F ↓ − F ↑ (1)

where αshort is surface total shortwave albedo, E is direct and diffuse shortwave irradiance on the
surface, F ↓ is atmospheric downward longwave flux, and F ↑ is total surface radiant exitance. Highly
developed areas, such as business and stadiums, should possess a lower NRB because the higher
albedos of bright construction materials typify them and therefore absorb less radiation than areas
with lower albedos, such as areas of medium development or vegetated spaces [38]. Also, urbanization
influences the distribution of heat fluxes related to NRB with a combination of drivers, including
replacement of vegetation with construction, reduced surface moisture, and the complexity of urban
morphometry [38]. BAEI is an expression of NRB, with low values for highly developed areas,
mid-ranged values for medium development, and high values for areas that are vegetated or comprised
of dark asphalt. Where index values increase over time, the increase is linked to the influence of human
beings on NRB, since NRB will rise for reasons linked to human development.
To validate this index as automatic, unsupervised classification of different grades of building
presence, an evaluation was made of the usefulness of the index to chronicle stages of urban growth
for Miami using the mean as a measure of central tendency, including the influence of vegetation.
Enhancements were derived for six satellite images in a time-series spanning 1985–2016. If thresholds
are not applied, BAEI rasters for Miami are highlighted by the formation of shadows cast by high-rise
infrastructure in downtown neighborhoods. The visualizations need to be refined by omitting the
formation of shadows to display an enhanced gradient visualization of an urban environment.
Histogram analysis is a method for identifying a statistical break in the range of index values that
will extract the target features from the rest of the data [39]. Thresholds were determined through a
trial-and-error process of raster histogram analysis performed in Imagine to establish the best visual fit
for separating shadows. BAEI rasters were geoprocessed with conditional thresholds for Landsat 5
and Landsat 8. Note that conditional thresholds of 2.5 and 2.4 were assigned to Landsat 5 and Landsat
8, respectively, to enhance visualizations. A different threshold for Landsat 8 is necessary because
it possesses different spectral bandwidths and cannot be directly cross-referenced with Landsat 5.
Cross-referencing can be performed with post-classification image comparison and post-classification
change detection.
A visual and statistical comparison was made of the time-series for the entire city. Due to the
apparent influence of urban canopy in the southern portion of the city on index values in the first
time-series, a shape truncated of the three southernmost neighborhoods was created. Rasters clipped
to the truncated area were necessary to evaluate how well this index tracks the accumulation of
constructed features over time as opposed to the accumulation of urban arbor. Further scrutiny was
given to the truncated area concerning the influence of vegetation.
NDVI has widely proven useful in classifying the presence and vigor of plants visible in remote
sensing data [8–10,29]. Index values are linear, −1 + 1, with values in the direction of 1 indicating the
presence of healthy vegetation.

NDVI = (NIR − RED)/(NIR + RED) (2)

NDVI was derived for a time-series of the truncated area for statistical and visual comparison.
Note that Miami-Dade County does not possess significant arbor compared to other urban areas in the
United States, having 20% tree cover in 2016 [40]. Zonal statistics for BAEI and NDVI were calculated
using the truncated area for a time-series of nine images spanning 1985–2016.
Remote Sens. 2020, 12, 2503 10 of 28

BAEI, utilized as a shadow index by not thresholding the raster, displays values for Miami with
a pronounced division between the shadows cast by high-rises and other urban area. Shadows are
focalized and further analyzed by establishing a threshold to separate cells that highlight shadows.
Subsequently, a time-series of urban land use classification maps incorporating the extracted shadow
highlights was created to depict a formation trend. A geosimulation incorporating a 2003 and
2008 image time-series in combination with directional land use polynomial transition trends as
predictors was performed to predict the construction of high-rise infrastructure for 2016 in the Brickell
neighborhood (see Section 2.7). Consideration was given regarding how NBI and BAEI may be
composited to take advantage of the useful properties of both enhancements.

2.6. Evaluating the Normalized Difference Concrete Condition Index for Incorporation in Urban Development
Intensity Classifications
It is possible to composite enhancements as an image stack, whereby the desirable properties
of each enhancement can be translated to a single output. NDCCI derived from the 2016 Landsat
8 data visualized Miami’s urban fabric as a network of streets containing pronounced areas of low
development, collections of homes, and varying stages of development in areas of business and dense
transport. Consideration was given to this index as the basis for an enhancement composite with
NBI and BAEI based on a comparison of zonal statistics with NDVI referencing urban land use types.
The composite was then classified to identify pronounced spectral groupings based on the feature
extraction achieved by combining enhancements. That is to say, differing states of urban intensity are
delineated according to their unique range within each built-up index.
NDCCI was developed by Samsudin et al. (2016) [20] with data from the Worldview-3 satellite
to assess the condition of concrete roofs in high-resolution data. Gu et al. (2018) [21] considered
the usefulness of the index for extracting building shapes from high-resolution data. Where NDVI
enhances the difference between near-infrared and red bands, this index compares the difference
between the green band. The index may be implemented to assess the presence and quality of
construction compared to the ability of NDVI to assess the presence and vigor of vegetation. Zonal
statistics were calculated using the FGDL data to evaluate the way NDCCI groups discrete urban types
compared to NDVI. Considering a reduced variance of spectral responses within discrete urban types
by NDCCI, the index is likely to be useful for identifying differences between urban land use types
with a classifier when composited with NBI and BAEI.
NBI orders the brightness of urban features from low-high, BAEI gives an order to NRB, and NDCCI
gives higher values according to the difference in spectral response between NIR and green wavelengths
(high for vegetated areas and lower according to the extent of development). Moreover, it is important
to visualize the differences between a classification performed with NDCCI and one performed with
NDVI to gain a better understanding of why it is advantageous to incorporate an index designed to
enhance construction compared to one designed to enhance vegetation. Thus, composites were created
for NDCCI/NBI/BAEI and NDVI/NBI/BAEI with the 2016 Landsat 8 data and analyzed with the Iso
Cluster unsupervised classifier by Ball & Hall (1965) [41].
According to Abbas et al. (2016) [24], iso cluster or ISODATA (iterative self-organizing data
analysis technique algorithm) is a novel iterative classifier that automates the process of feature
extraction, with the caveat that accuracy is based on a computer analysis of average values instead of
user-defined training samples necessary for supervised classification. The unsupervised classification
begins by assigning classes to arbitrary values that display statistical clustering. Then, every pixel
is assigned to the nearest cluster, the mean of each cluster is calculated, and a series of iterations is
performed until either the percentage of pixels that changes between iterations is too small or the
distances between vectorized means in a feature space is too small [24].
Both composites were classified using the ArcGIS default settings. The default pixel-based iso
cluster settings are maximum number of classes: 5, maximum number of iterations: 20, maximum
number of cluster merges per iteration: 5, maximum merge distance: 0.5, minimum samples per cluster:
Remote Sens. 2020, 12, 2503 11 of 28

20, and skip factor: 10 (referring to the number of x and y values to be skipped during class seeding).
These settings were not changed. The classification output was visually assessed and reclassified to
establish an interpretation of low-to-high urban development, in terms of arbitrary classes named
Urban 1-5. An unsupervised classification may serve as support data for a more refined supervised
classification. It must be noted that the classifications performed here and henceforward were done to
make comparisons with and improve upon NLCD 2016 urban intensity data, and the enhancements
were conditionally processed to only include cells classified as urban by the USGS. That classification
contains errors in the study area, where some vegetated areas are incorrectly classified as non-urban.
Based on a visual assessment of the unsupervised NDCCI/NBI/BAEI classification, classifying
urban development in terms of low, medium, and high is straightforward, and further analysis was
performed to create a refined classification with the SVM classifier by Cortes & Vapnik (1995) [42].
Pal & Mather (2005) [26] report success using SVM to classify Landsat 7 Enhanced Thematic Mapper
Plus data for Littleport, England into eight cover types: wheat, water, dry salt lake, hydrophytic
vegetation, vineyards, bare, pasture, and built-up. SVM is preferable to other classifiers, such as GML,
due to its ability to provide more accurate results and because it works well with a small number
of training samples [43]. The classifier draws upon statistical learning theory to identify decision
boundaries to separate classes with an optimal hyperplane in feature space [26,42]. For linearly
separable classes, it will identify the decision boundary that minimizes generalization error: the one
that leaves the greatest margin between the points, or “support vectors” of those classes closest to the
hyperplane [26,42]. Classes that cannot be linearly separated are handled the same way, though the
classifier performs a minimizing function handling the proportion of cells classified incorrectly [26,42].
Chen & Stow (2002) [44] affirm that the selection of several small polygons to serve as training
samples for each class, as opposed to a single polygon or a group of individual pixels, will maximize
time efficiency as well as accuracy when dealing with heterogenous features, such as urban land
use types, in moderate resolution imagery. Therefore, the training samples used here should be
polygons groups that capture all of the spectral properties of each intended class to avoid inheriting
spatial autocorrelation [44]. Spatial autocorrelation among samples would unfavorably homogenize
the statistical variance of each class, resulting in classes less representative of feature structure [44].
The NDCCI/NBI/BAEI composite was classified with pixel-based SVM in terms of arbitrary classes using
training samples selected by comparing areas of the composite to high-resolution basemap imagery
for 2016 available through ArcGIS in cross-reference with the FDEP data. Training samples were
established to separate open space, residential and vegetated residential, and business and transport
areas into classes Urban 1-3, indicating the intensity of development. SVM is a non-parametric classifier,
not assuming any statistical distribution of samples, and uneven samples are selected since this strategy
is reported to increase classification accuracy [43] (see Table 3). The default analysis setting of maximum
number of samples per class: 500, which may influence the accuracy of results, was not changed [43].
Area statistics from the FDEP data were calculated to assess features typifying each class. It must
be noted that the FDEP data covers Miami without breaks between features, which is advantageous
since the SVM classification should correctly aggregate the entire city. NLCD Percent Developed
Imperviousness data can be conditionally geoprocessed to select areas possessing a certain intensity of
development as defined by the USGS. That selection might then be fused with this classification to
maximize the amount of detail regarding urban intensity derived from a satellite image.

Table 3. Breakdown of the number of training polygons and pixels sampled for each SVM class.

Class Polygons Pixels

Urban 1 6 931
Urban 2 9 3572
Urban 3 9 1580
Total 24 6083
Remote Sens. 2020, 12, 2503 12 of 28

USGS NLCD urban intensities possess shortcomings such as fragments of open space incorrectly
classified as low intensity, large patches of low intensity incorrectly classified as open space due to
vegetation, high intensity areas that are saturated and overflow into areas of lower intensity, and areas
of high development incorrectly classified into the lower classes. Referring to Figure 3, the NLCD122016
Remote Sens. 2020, 12, x FOR PEER REVIEW of 29
urban intensity map of Miami displays numerous noticeable errors. Four examples are circled in yellow
and numbered:
circled in yellow(1) fragments
and numbered: of Mount Nebo and
1) fragments Memorial
of Mount Plan
Nebo andFlagler Memorial
Memorial Parks incorrectly
Plan Flagler Memorial
classified as low intensity, (2) a vegetated patch of the Coconut Grove neighborhoods
Parks incorrectly classified as low intensity, 2) a vegetated patch of the Coconut Grove neighborhoods incorrectly
classified
incorrectlyasclassified
open space, as (3)
openan space,
overflow of high
3) an intensity
overflow along
of high a commercial
intensity along astrip in the Little
commercial Havana
strip in the
neighborhood, (4) Marlins Park incorrectly classified in lower classes. The
Little Havana neighborhood, 4) Marlins Park incorrectly classified in lower classes. The SVM SVM classification was
conditionally geoprocessed to increase detail by replacing any cells with an imperviousness
classification was conditionally geoprocessed to increase detail by replacing any cells with an between
50% and 79% with
imperviousness a new class.
between 50% and The79%
result of the
with SVM/percent
a new class. Thedeveloped imperviousness
result of the SVM/percent fusion still
developed
displayed an incorrect classification for Marlins Park. It was revised with a conditional
imperviousness fusion still displayed an incorrect classification for Marlins Park. It was revised with methodology
to implement methodology
a conditional the iso clusterto classification
implement as theaiso
correction.
cluster classification as a correction.

Figure 3. USGS NLCD 2016 urban intensities with noticeable errors denoted.

The SVM/Percent Developed Imperviousness fusion was conditionally geoprocessed to replace

any cells classified as Urban 5 in the iso cluster
cluster classification
classification with a new class. To finally amend the
NLCD urban intensity classification, the fifth class class was
was decomposed
decomposed into into the
the fourth.
fourth. Area statistics
statistics
were calculated
calculated from the
the FDEP
FDEP data
data for
for the
the five-class
five-class and
and four-class
four-class maps
maps toto assess
assess features
features that
typify each
each class.
class.Maps
Maps created
created forfor
thisthis
partpart of study
of the the study
werewere projected
projected to theto
USAtheContiguous
USA Contiguous
Albers
Albers EqualConic
Equal Area Area Conic
USGSUSGS coordinate
coordinate system
system andand resampled
resampled with
with cubicconvolution
cubic convolutiontoto match
match the
original NLCD products, which are distributed as cubic convolution convolution maps
maps in the Albers projection.
projection.
The semi-automated
semi-automated workflow
workflowfor forcreating
creatingnewnew urban
urban intensity
intensity maps
maps is summarized
is summarized in Figure
in Figure 4. The4.
The
nextnext section
section willwill describe
describe thethe creation
creation of of a geosimulationofofthe
a geosimulation theformation
formationof ofhigh-rise
high-rise infrastructure
infrastructure
in the Brickell neighborhood
neighborhood incorporating
incorporating BAEIBAEI asas aa shadow
shadow index.
index.
Remote Sens. 2020, 12, 2503 13 of 28
Remote Sens. 2020, 12, x FOR PEER REVIEW 13 of 29

Figure 4. Urban
Figure 4. Urban intensity classification workflow
intensity classification workflow diagram.
diagram.

2.7. Simulating the Development of High-Rise Infrastructure with Land Change Modeler
2.7. Simulating the Development of High-Rise Infrastructure with Land Change Modeler
An ANN is a dynamic computerized model that attempts to duplicate human actions and learning
An ANN is a dynamic computerized model that attempts to duplicate human actions and
functions. MLP is the most commonly implemented ANN. A backpropagation algorithm, learning
learning functions. MLP is the most commonly implemented ANN. A backpropagation algorithm,
based on gradient descent when the network informs itself of a prediction error, is useful for predicting
learning based on gradient descent when the network informs itself of a prediction error, is useful
complex values [45]. The challenges of operating MLP ANN include configuring its parameters and
for predicting complex values [45]. The challenges of operating MLP ANN include configuring its
establishing the correct inputs used to calculate a prediction [45].
parameters and establishing the correct inputs used to calculate a prediction [45].
The Markov chain is an empirical method of land change modeling implemented to generate a
The Markov chain is an empirical method of land change modeling implemented to generate a
transition probability matrix that serves as a basis for space-time-series analysis. A first-order Markov
transition probability matrix that serves as a basis for space-time-series analysis. A first-order Markov
chain
chainmodel
model isis defined
defined as as aa chain,
chain, where
where the the probability
probability of of the
the future
future state
state depends
dependsonly onlyon onthe
the
present
presentstate
stateand
andnot notononpreceding
precedingstates,
states, and
and every
every point
point in time is
in time is the
the next
next step.
step.ItItisisappropriate
appropriatetoto
incorporate
incorporate this technology in modeling urban systems because the transformation properties ofofland
this technology in modeling urban systems because the transformation properties land
change have to do with predictable transitions and
change have to do with predictable transitions and steady states [2]. steady states [2].
The
TheLCM LCMmodule
module of of TerrSet
TerrSet cancan function
function as as aa geostatistical Markov chain
geostatistical Markov chain predictor,
predictor,utilizing
utilizing
the functionality of MLP ANN. The module accepts input in the form
the functionality of MLP ANN. The module accepts input in the form of classified before-and-after of classified before-and-after
rasters
rastersas aswell
wellas aspredictor
predictorvariables.
variables. Typical
Typical geographic predictors of
geographic predictors of urban
urban development
developmentinclude include
elevation,
elevation,distance
distance to to transportation,
transportation, population,
population, and and income.
income. In Inaddition
additiontotothese
theseusual
usualpredictors,
predictors,
polynomial transition trends can be generated within the module from
polynomial transition trends can be generated within the module from the before-and-after data the before-and-after dataused
used
totointerpolate
interpolatethe theprediction.
prediction. Although geosimulations may
Although geosimulations may have
havedifficulty
difficultypredicting
predictingthe theoccurrence
occurrence
ofofisolated
isolatedfeatures
features as as they
they appear
appear over
over time, a novel method method for for predicting
predictingthe thegrowth
growthof ofdiscrete
discrete
urban features is evaluated
urban features is evaluated [1–3]. [1–3].
BAEI
BAEIderived
calculated from Landsat
using 5 data
Landsat for 2003
5 data and 2008,
for 2003 and Landsat
and 2008, and Landsat 8 data for 2016
8 data was clipped
for 2016 to the
was clipped
Brickell
to the neighborhood.
Brickell neighborhood.Without aWithout
conditional threshold, the
a conditional shadows
threshold, thecast by high-rise
shadows cast infrastructure
by high-rise
were rendered clearly
infrastructure in higherclearly
were rendered index values.
in higher The output
index conditional
values. The outputrasters were clipped
conditional to anwere
rasters area
ofclipped
interesttowherean areashadows seem to
of interest be developing
where shadows over seema totime-series
be developingfor visualization. By focalizing
over a time-series for
the BAEI rasters
visualization. Bywith a one-cell
focalizing theradius, the formation
BAEI rasters of shadows
with a one-cell wasthe
radius, highlighted
formationby ofashadows
group of washigh
highlighted
values. Eachby a group of high
focalization values.
was then Each focalization
conditionally was thenthrough
thresholded, conditionally thresholded,
a process through
of trial-and-error
a process
based of trial-and-error
strictly based strictly
on visual interpretation to on visual interpretation
establish the best fit. The to establish
value 1 thewasbest visual fit.
assigned The
to areas
value
with 1 was
high assigned
values and 0totoareas
areas with
withhigh
lowvalues
values. and 0 tooutput
The areas with low values.
conditional rasterThe output
was conditional
symbolized with
raster was symbolized with transparency over rasters symbolized
transparency over rasters symbolized with a standard deviation stretch and gamma adjustments. with a standard deviation stretch
and gamma
Upon adjustments.itUpon
visual inspection, visual inspection,
was assumed it was assumed
that attempting to predict that attempting
the to predict
area highlighted in the
thearea
2016
highlighted
data, in theto2016
by referring data,and
the 2003 by referring
2008 images,to thewould
2003 and 2008approached
be best images, would be best approached
by attempting to generalizeby
attempting to generalize the
the prediction by “circling” the area. prediction by “circling” the area.
Remote Sens. 2020, 12, 2503 14 of 28

For the 2003 and 2008 data, manual classifications were performed for the Brickell neighborhood,
assigning each 30 × 30 m cell to either Road, Low Development, or Business classes. Class assignment
was attributed to the polygons of a 30 × 30 m vector fishnet which was rasterized and geoprocessed
to include the shadows that overlap business areas. Conditional focalizations created for Brickell
were integrated into the manual classifications to create maps depicting the formation of shadows
for each year in the time-series. These maps were used as LCM inputs. From the input maps,
the module can automatically calculate polynomial transition trends for use as predictor gradients
in the geosimulation. During the model fitting routine, where the predictive capacity of every
combination of polynomial trends can be examined, 2nd and 3rd order polynomial transition trend
maps were selected as best fit predictors for the visualization. Only polynomial transition trends
were incorporated as predictors in this model. Default parameter settings, such as software-generated
transition probabilities, were used when running the geosimulation. Because this module features
many settings, some of high complexity, it is beyond the scope of this paper to discuss further the
configuration of the geosimulation. Nonetheless, the research conducted herein should be reproducible
given the simplicity of the novel method.

3. Results

3.1. Objected-Oriented Urban Feature Extraction with the New Built-Up Index
When visualized, high NBI values are found in areas of pronounced development, such as
downtown neighborhoods and the industrial strips in the Allapattah neighborhood to the northwest
and in the Little Haiti neighborhood to the north. Open areas checkered throughout, such as golf
courses and cemeteries, return low values. Both of those area types are well contrasted by areas of
predominantly residential units possessing mid-ranged values along their borders. Zonal statistics
were calculated for discrete urban land use types for NBI derived from the 2016 Landsat 8 data
(see Table 4). Statistically, there appeared to be pronounced differences in central tendency between
the types listed. The mean values range from high, attributable to industrial land, down to commercial
and institutional, down again to residential, and then recreational. The index arranges the brightness
of urban surface features in linear order.

Table 4. Zonal statistics of NBI derived from 22 October 2016 Landsat 8 OLI data, referencing FGDL data.

Land Use Area (sq km) Min Max Range Mean Std
Centrally Assessed 1 0.33 0.0195 0.2788 0.2593 0.1262 0.0535
Industrial 1.97 0.0209 0.5135 0.4927 0.1698 0.0663
Institutional 2 2.51 0.0002 0.4243 0.4241 0.0846 0.0639
Public/Semi-public 3 7.99 0.0004 0.7474 0.7470 0.0931 0.0711
Recreation 0.6 0.0037 0.3645 0.3607 0.0631 0.0564
Residential 36.78 0.0004 0.4011 0.4008 0.0785 0.0421
Retail/Office 7.52 0.0044 0.7848 0.7804 0.1356 0.0701
Vacant Non-residential 2.61 0.0056 0.4932 0.4877 0.1120 0.0754
Vacant Residential 1.84 0.0035 0.2860 0.2825 0.0607 0.0449
1Centrally assessed refers to, for instance, utilities. 2 Instutional refers to religious, educational, social, cultural,
or healthcare. 3 Public/Semi-public refers to schools, governmental.

By utilizing the mean shift between urban land use types, NBI rasters of urban areas can
be segmented into object primitives. Primitives emphasizing brightness, shape, and texture were
visualized in Figure 5, where the scaled values represent the brightness captured by each segment.
Visual assessment of the primitives yields the idea that urban land use types can be identified as
distinct objects. The properties of a segment extracted from this index should be cross-referenced with
ground truth data to verify the usefulness of this methodology.
 spatial detail of 15, and a minimum segment size of 1000 (see Figure 6). The feature was selected
based on a visual assessment of the brightness and configuration of surface features. The object is
typified by the presence of industrial and commercial land use (see Table 5). The results indicate that,
through OBIA, it is possible to create maps that display urban land use as a series of classified
geometries. Such methodology may entail automated geoprocessing of a large series of object
Remotesettings,
Sens. 2020,and the output rasters can then be reviewed by an operator to select the most fitting
12, 2503 15 of 28
segments as shapes to map the target urban area.

FigureFigure
5. NBI5. NBI object
object primitives
primitives derivedfrom
derived from Landsat
Landsat 88OLIOLI10/22/16 data; 2016
22 October (a) small segment
data; size segment
(a) small
emphasizing surface
size emphasizing surface brightness
brightness (spectral detail:
(spectral 13, spatial
detail: detail: 8,
13, spatial minimum
detail: segment size
8, minimum in
segment size in
pixels:pixels: 1),moderate
1), (b) (b) moderate segment
segment size
size emphasizingsurface
emphasizing surface objects
objects (spectral
(spectraldetail: 15,15,
detail: spatial detail:
spatial detail: 10,
10, minimum
minimum segmentsegment
size in size in pixels:
pixels: 5), (c)5), (c) large
large segment
segment sizesize emphasizing
emphasizing surface
surface texture(spectral
texture (spectraldetail:
detail: 17, spatial detail: 12, minimum segment size in pixels: 10).
17, spatial detail: 12, minimum segment size in pixels: 10).

A single object was extracted by vectorizing a segment created with a spectral detail of 15.5,
a spatial detail of 15, and a minimum segment size of 1000 (see Figure 6). The feature was selected
based on a visual assessment of the brightness and configuration of surface features. The object is
typified by the presence of industrial and commercial land use (see Table 5). The results indicate
that, through OBIA, it is possible to create maps that display urban land use as a series of classified
geometries. Such methodology may entail automated geoprocessing of a large series of object settings,
and the output rasters can then be reviewed by an operator to select the most fitting segments as shapes
to map the target urban area.
Remote Sens. 2020, 12, x FOR PEER REVIEW 16 of 29

Figure 6. NBI-derived object of industrialized land in Little Haiti.

Figure 6. NBI-derived object of industrialized land in Little Haiti.
Table 5. Area statistics of NBI-derived object, referencing FGDL data.

Land Use Area (sq km)

Centrally Assessed 0.06
Industrial 0.43
Institutional 0.05
Public/Semi-public 0.04
Remote Sens. 2020, 12, 2503 16 of 28

Table 5. Area statistics of NBI-derived object, referencing FGDL data.

Land Use Area (sq km)

Centrally Assessed 0.06
Industrial 0.43
Institutional 0.05
Public/Semi-public 0.04
Residential 0.03
Retail/Office 0.14
Vacant Nonresidential 0.05

3.2. Quantifying Infill Development with the Built-Up Area Extraction Index
Visual and statistical analysis was made of a BAEI time-series, including data from Landsat
5 and Landsat 8, for the entire city (see Figure 7). The time-series corresponds with a population
influx and subsequent accumulation of low-albedo construction materials due to human development.
ThereRemote
appears to be a trend of increasing central tendency for the Landsat 5 data. The NRB appears
Sens. 2020, 12, x FOR PEER REVIEW 17 of 29
to be
increasing more rapidly over time, given exponential population growth and subsequent development
tendency
trends occurring of BAEI
morefor areas not
recently. shrouded
Despite by the timeframe,
a smaller urban canopy.theMoreover, zonal statistics
change between 2003 and were 2011 is
morecalculated
pronounced for the
thantruncated
the changearea between
with a time-series
1985 andof1998.
nine images spanning 1985–2016,
Image comparison between including
the Landsat
data
8 data froman
shows Landsat
increase 5 and
in 8. Table tendency
central 6 displays despite
a trend ofthe
increasing BAEI values
small, two-year timecorresponding
frame between withthem.
decreasing NDVI values, indicating a negative correlation. The results suggests that BAEI is useful
Evaluating the entire time-series, areas towards the coast and within other highly developed places
for measuring the extent of large-scale urbanization as a low-high gradient, while NDVI is useful for
have a higher NRB compared to areas farther inland.
measuring the vigor of plants.

Figure 7. BAEI change detection time-series, giving mean index values, derived from Landsat data;
Figure 7. BAEI change detection time-series, giving mean index values, derived from Landsat data;
(a) 5 TM 11/02/85, (b) 11/06/98, (c) 11/20/03, (d) 11/10/11, (e) 8 OLI 10/17/14, (f) 10/22/16.
(a) 5 TM 2 November 1985, (b) 6 November 1998, (c) 20 November 2003, (d) 10 November 2011, (e) 8 OLI
17 October 2014, (f) 22 October 2016.
Remote Sens. 2020, 12, 2503 17 of 28

A similar time-series analysis, comparing BAEI to NDVI, was conducted with the Landsat 5 data
for part of Miami truncated of neighborhoods with developed urban canopies to minimize the influence
of vegetation on BAEI returns (see Figure 8). Visual comparison with NDVI indicates that BAEI is
more useful for tracking shifts in urban morphology. The BAEI time-series possesses an increasing
mean, while the NDVI mean is decreasing. There is a clear, continual increase in the central tendency
of BAEI for areas not shrouded by the urban canopy. Moreover, zonal statistics were calculated for the
truncated area with a time-series of nine images spanning 1985–2016, including data from Landsat 5
and 8. Table 6 displays a trend of increasing BAEI values corresponding with decreasing NDVI values,
indicating a negative correlation. The results suggests that BAEI is useful for measuring the extent of
large-scale urbanization as a low-high gradient, while NDVI is useful for measuring the vigor of plants.
Remote Sens. 2020, 12, x FOR PEER REVIEW 18 of 29

Figure 8. Change detection time-series, giving mean index values, derived from Landsat 5 TM data;
Figure 8. Change detection time-series, giving mean index values, derived from Landsat 5 TM data;
(a) 11/02/85 BAEI, (b) 11/06/98 BAEI, (c) 11/10/11 BAEI, (d) 11/02/85 NDVI, (e) 11/06/98 NDVI, (f)
(a) 2 November 1985 BAEI, (b) 6 November 1998 BAEI, (c) 10 November 2011 BAEI, (d) 2 November
11/10/11 NDVI.
1985 NDVI, (e) 6 November 1998 NDVI, (f) 10 November 2011 NDVI.

TableTable 6. Zonal
6. Zonal statistics
statistics of BAEI
of BAEI andandNDVI
NDVIvalues
values derived
derived from
fromLandsat
Landsat5 TM (1985-2011)
5 TM and 8and
(1985–2011) OLI8 OLI
data (2014 & 2016), referencing the truncated study area.
data (2014 & 2016), referencing the truncated study area.
Year Mean BAEI Mean NDVI
1985 Year Mean BAEI
1.283 Mean NDVI0.3388
1998 1985 1.283
1.304 0.3388 0.3121
1999 1998 1.304
1.284 0.3121 0.3142
2003 1999 1.284
1.332 0.3142 0.3074
2008 2003 1.332
1.364 0.3074 0.2945
2009 2008 1.364
1.299 0.2945 0.2891
2009 1.299 0.2891
2011 1.377 0.3049
2011 1.377 0.3049
2014 1.353 0.3900
2014 1.353 0.3900
2016 2016 1.367
1.367 0.3897 0.3897

3.3. New Urban Intensities Derived from the Normalized Difference Concrete Condition Index
Building on the measurements achieved with NBI and BAEI, enhancements can be composited
into a stack that could be utilized for the creation of discrete urban classes [27]. Miami’s urban fabric
is rendered with NDCCI as a pronounced network of streets within open space, residential, and
business areas. An effective enhancement composite would emphasize differences that exist between
features within an urban landscape, where NBI, BAEI, and NDCCI do not yield similar renderings.
Since NDVI is useful for identifying areas with or without vegetation, zonal statistics were calculated
Remote Sens. 2020, 12, 2503 18 of 28

3.3. New Urban Intensities Derived from the Normalized Difference Concrete Condition Index
Building on the measurements achieved with NBI and BAEI, enhancements can be composited
into a stack that could be utilized for the creation of discrete urban classes [27]. Miami’s urban fabric is
rendered with NDCCI as a pronounced network of streets within open space, residential, and business
areas. An effective enhancement composite would emphasize differences that exist between features
within an urban landscape, where NBI, BAEI, and NDCCI do not yield similar renderings. Since NDVI
is useful for identifying areas with or without vegetation, zonal statistics were calculated from urban
land use types for both NDCCI and NDVI, both derived from the 2016 Landsat 8 data. NDCCI
possesses less variation within classes, according to the coefficient of variation (CV) statistic, compared
to NDVI (see Table 7). CV is calculated by dividing the standard deviation by the mean and may be
written as a percentage; it gives a measure of the amount of statistical variation that exists between
classes, allowing for direct comparison.

Table 7. Zonal statistics of 22 October 2016 NDCCI and NDVI, referencing FGDL data.

NDCCI
Land Use Area (sq km) Min Max Range Mean Std CV
Centrally Assessed 0.33 0.004 0.701 0.698 0.345 0.124 35.96%
Industrial 1.97 −0.100 0.649 0.749 0.221 0.106 47.96%
Institutional 2.51 −0.316 0.881 1.197 0.461 0.190 41.12%
Public/Semi-public 7.99 −0.136 0.914 1.050 0.439 0.189 42.93%
Recreation 0.6 −0.177 0.824 1.002 0.515 0.208 40.33%
Residential 36.78 −0.289 0.915 1.205 0.481 0.138 28.61%
Retail/Office 7.52 −0.811 0.886 1.697 0.262 0.141 53.98%
Vacant Non-residential 2.61 −0.462 0.827 1.289 0.392 0.186 47.42%
Vacant Residential 1.84 −0.185 0.845 1.030 0.535 0.172 32.18%
NDVI
Land Use Area (sq km) Min Max Range Mean Std CV
Centrally Assessed 0.33 0.014 0.785 0.771 0.322 0.148 45.80%
Industrial 1.97 −0.065 0.699 0.764 0.185 0.112 60.48%
Institutional 2.51 −0.217 0.955 1.171 0.463 0.229 49.45%
Public/Semi-public 7.99 −0.057 0.972 1.029 0.442 0.225 51.03%
Recreation 0.6 −0.081 0.904 0.985 0.536 0.247 46%
Residential 36.78 −0.123 0.958 1.081 0.461 0.162 35.21%
Retail/Office 7.52 −0.633 0.944 1.576 0.234 0.153 65.41%
Vacant Non-residential 2.61 −0.296 0.905 1.201 0.384 0.220 57.31%
Vacant Residential 1.84 −0.117 0.925 1.042 0.550 0.207 37.60%

Iso cluster classifications for NDCCI/NBI/BAEI and NDVI/NBI/BAEI composites were created for
visual comparison. Both composites were assigned arbitrary classes of Urban 1-5, yielding a view of
development intensities based on respective enhancements. Areas in classes 1–3 are typically open
space, vegetated, or residential, whereas areas of intense development are found in classes 4 and 5.
The maps were compared for the clarity of discrete urban features. Examining the two maps in Figure 9,
finer surface features such as Marlins Park, highways, and streets are more vividly rendered with the
NDCCI/NBI/BAEI classification. The features in the NDVI/NBI/BAEI are comparatively saturated,
with the compromise of less detail. Unsupervised classification can yields maps that, by themselves,
may be valuable for the identification of specific surface features and for vectorizing those features into
point, polygon, or line shapes. Therefore, further analysis was conducted with a supervised classifier
to map the spatial configuration of Miami’s urban features more accurately.
Remote Sens. 2020, 12, 2503 19 of 28
Remote Sens. 2020, 12, x FOR PEER REVIEW 20 of 29

Figure 9.
Figure 9. IsoIso
cluster classifications
cluster of urban
classifications land use
of urban intensity
land based on spectral
use intensity based on index composites
spectral index
derived from 10/22/16 Landsat 8 OLI data; (a) NDCCI/NBI/BAEI composite, (b) NDVI/NBI/BAEI
composites derived from 22 October 2016 Landsat 8 OLI data; (a) NDCCI/NBI/BAEI composite,
composite.
(b) NDVI/NBI/BAEI composite.

3.3.1.
3.3.1.Classifying
ClassifyingUrban
Urban Intensities
Intensities with
with Support
Support Vector
Vector Machine
Machine
AAsupervised
supervisedclassification
classification waswas created
created with
with the SVM classifier to to build
build upon
upon thetheunsupervised
unsupervised
classification
classification ofof the NDCCI/NBI/BAEI composite. Upon examining Figure
NDCCI/NBI/BAEI composite. Upon examining Figure 10a, it is clear that 10a, it is clearthe
that
SVM the
SVM
method method identifies
identifies urbanurban
intensityintensity
in threeinsimple
three simple classes
classes and and features
features corrections corrections to class
to class conflicts
conflicts
found infound
NLCDin2016.NLCD 2016.
Zonal Zonal statistics
statistics were calculated
were calculated for urbanfor landurban land use
use types typesthe
to assess to assess
accuracythe
accuracy of the classification. Referring to Table 8, Urban 1 is typified by recreation
of the classification. Referring to Table 8, Urban 1 is typified by recreation and other low intensity and other low
intensity development,
development, Urban 2Urban 2 by residential
by residential areas,areas, and Urban
and Urban 3 by3 by business
business andand otherhigh
other highintensity
intensity
development.
development.Capitalizing
Capitalizingonon thethe
spectral
spectralconfiguration
configuration initially identified
initially for NBI
identified for in
NBISection 3.1, highly
in Section 3.1,
highly developed
developed areas areareas are distinguished
distinguished from open from open
areas, areas,
clearly clearly bordered
bordered by areas of bymostly
areas residential
of mostly
residential development.
development. Where NLCD Where
featuresNLCD features an
an additional additional
class to describe class to describe
medium medium
intensity, furtherintensity,
analysis
further
was analysis
conducted towas
fuseconducted to fuse
that class with thethat
SVM class with the SVM
classification classification
to increase to increase
cartographic cartographic
detail. The results
detail.
infer The
this is aresults infer
globally this is a globally
replicable method replicable
that can be method that globally
replicated can be replicated
to provide globally
urbantoareaprovide
maps
urban
that area maps
describe that describe
development development
in three classes. in three classes.

Table8.8.Area
Table Areastatistics
statisticsfor
forSVM
SVMclassification
classificationofofNDCCI/NBI/BAEI
NDCCI/NBI/BAEIcomposite
compositederived
derivedfrom
from2210/22/16
October
Landsat 8 OLI data, referencing FDEP data.
2016 Landsat 8 OLI data, referencing FDEP data.
Land UseLand Use Urban
Urban 1 (sq1 km)
(sq km) Urban 2 Urban
Urban2 3 Urban 3
Commercial and Services 0.630 3.480 10.886
Commercial and Services 0.630 3.480 10.886
Industrial 0.027 0.460 2.651
Industrial 0.027 0.460 2.651
Institutional 0.390 2.156 1.989
Institutional 0.390 2.156 1.989
Open Land 0.054 0.070 0.098
Open Land 0.054 0.070 0.098
Recreational
Recreational 0.7360.736 0.671 0.671
0.494 0.494
Residential High Density
Residential High Density 0.2600.260 5.098 5.0984.091 4.091
Residential Low Density
Residential Low Density 0.0100.010 0.065 0.065
0.000 0.000
Residential Medium
Residential DensityDensity
Medium 1.3261.326 40.167 40.167
5.099 5.099
Transportation
Transportation 0.0490.049 0.854 0.8542.095 2.095
UtilitiesUtilities 0.0040.004 0.044 0.0440.138 0.138
Remote Sens. 2020, 12, 2503 20 of 28
Remote Sens. 2020, 12, x FOR PEER REVIEW 21 of 29

3.3.2.The
3.3.2. TheFusion
Fusion Map
Map
USGSclassifies
USGS classifies areas
areas identified
identified as
as possessing
possessing 50-79%
50–79%impervious
impervioussurface
surfaceasasmedium
mediumintensity,
intensity,
and the class is rendered as a detail of streets and well-developed buildings in theNLCD
and the class is rendered as a detail of streets and well-developed buildings in the NLCDclassified
classified
product.The
product. The SVM
SVM classification
classification reduced
reduced thethe errors
errors related
relatedto
tofeature
featurestructure
structurefound
foundininNLCD
NLCD2016.
2016.
InInreciprocal,
reciprocal,areas
areas identified
identified by USGS
by USGS as medium
as medium intensity
intensity were were conditionally
conditionally geoprocessed
geoprocessed to
to overlay
overlay the SVM classification as a new third class typified by moderate intensity business and
the SVM classification as a new third class typified by moderate intensity business and medium-to-high
medium-to-high intensity residential development, yielding classes Urban 1–4 (see Figure 10b,c).
intensity residential development, yielding classes Urban 1–4 (see Figure 10b,c). Upon visual inspection,
Upon visual inspection, it was discerned that some well-developed features, such as Marlins Park,
it was discerned that some well-developed features, such as Marlins Park, may still be classified
may still be classified incorrectly into lower classes after fusing the SVM classification with the NLCD
incorrectly into lower classes after fusing the SVM classification with the NLCD percent developed
percent developed imperviousness data. To correct those features, the fifth class of the unsupervised
imperviousness data. To correct those features, the fifth class of the unsupervised NDCCI/NBI/BAEI
NDCCI/NBI/BAEI classification was overlaid as a new class. Upon subsequent visual inspection, the
classification was overlaid as a new class. Upon subsequent visual inspection, the result of adding this
result of adding this fifth class was a successful fusion. the brightest features, such as Marlins Park,
fifth class was a successful fusion. the brightest features, such as Marlins Park, freeways, and high
freeways, and high intensity industrial and commercial development were classified as Urban 5. By
intensity industrial and commercial development were classified as Urban 5. By decomposing the fifth
decomposing the fifth class into the fourth, a new four-class urban intensity map without the
class into the fourth, a new four-class urban intensity map without the pronounced errors of NLCD
pronounced errors of NLCD 2016 was created. The five-class fusion map does much to divide areas
2016 was created. The five-class fusion map does much to divide areas of the highest intensity from the
of the highest intensity from the fourth class, and the visualization may possess an aesthetic appeal.
fourth class, and the visualization may possess an aesthetic appeal. The four-class fusion map is more
The four-class fusion map is more homogenous compared to NLCD 2016, having classification errors
homogenous compared to NLCD 2016, having classification errors associated with urban intensities
associated with urban intensities derived only from the percent developed imperviousness corrected
derived only10d–f).
(see Figure from the percent developed imperviousness corrected (see Figure 10d–f).

Figure10.
Figure 10.2016
2016 Miami
Miami urban
urban land land use classifications
use classifications derivedderived from 8Landsat
from Landsat OLI data; 8 (a)
OLI data; (a)
NDCCI/NBI/
NDCCI/NBI/BAEI
BAEI composite SVM, composite
(b) NLCDSVM,
2016 (b) NLCD
Percent 2016 Percent
Developed Developed Imperviousness,
Imperviousness, (c) fusion of
(c) fusion of NDCCI/NBI/BAEI
NDCCI/NBI/BAEI
SVM SVM and Imperviousness,
and Percent Developed Percent Developed (d)Imperviousness,
NDCCI/NBI/BAEI (d)composite
NDCCI/NBI/BAEI composite
iso cluster, iso
(e) five-class
cluster, (e) five-class fusion of Percent Developed Imperviousness with NDCCI/NBI/BAEI
fusion of Percent Developed Imperviousness with NDCCI/NBI/BAEI SVM and iso cluster, (f) four-class SVM and
iso cluster,
fusion (f) four-class
of Percent Developedfusion of Percent Developed
Imperviousness Imperviousness
with NDCCI/NBI/BAEI with
SVM NDCCI/NBI/BAEI
and iso cluster. SVM
and iso cluster.
Remote Sens. 2020, 12, 2503 21 of 28

Zonal statistics were calculated for the NLCD 2016 Miami urban intensities as well as the
NDCCI/NBI/BAEI five-class and four-class fusion maps. Improvements were made in the fusion
classifications (see Table 9). For example, NLCD 2016 open space is typified by both open areas and
urban canopy, and the refined fusion classification effectively separates open areas from the urban
canopy. The distribution of urban intensity classes among land use types is similar between the
four-class maps. However, NLCD 2016 shows a large amount of residential areas classified incorrectly
in the first class and noticeably less area in the fourth class. Based on the results, it can be concluded
this is an effective methodology for producing urban intensity maps that correctly delineate detailed
stages of development according to enhanced reflectance retrieved from a remote sensor. It is proposed
that the process of classifying urban intensity with a satellite image in this analysis is effective for
creating cartographic products of urban intensity that are superior to those currently found in NLCD.

Table 9. Area statistics of 2016 urban intensity classifications, referencing FDEP data.

USGS NLCD 2016

Land Use Open Space (sq km) Low Medium High
Commercial and Services 0.546 1.434 5.171 7.932
Industrial 0.008 0.109 0.663 2.367
Institutional 0.457 1.047 1.932 1.112
Open Land 0.033 0.094 0.077 0.022
Recreational 0.795 0.504 0.434 0.179
Residential High Density 0.320 2.218 4.984 1.967
Residential Low Density 0.043 0.032 0.002 0.000
Residential Medium
5.084 22.002 17.768 1.861
Density
Transportation 0.111 0.444 1.179 1.288
Utilities 0.000 0.017 0.055 0.113
Five-class 2016 % developed imperviousness, NDCCI/NBI/BAEI SVM and iso cluster fusion
Land Use Urban 1 (sq km) Urban 2 Urban 3 Urban 4 Urban 5
Commercial and Services 0.560 1.694 4.271 4.210 4.163
Industrial 0.020 0.252 0.502 0.916 1.440
Institutional 0.330 1.162 1.645 0.630 0.717
Open Land 0.048 0.041 0.058 0.028 0.046
Recreational 0.669 0.446 0.330 0.170 0.206
Residential High Density 0.209 2.221 4.577 1.434 0.970
Residential Low Density 0.009 0.059 0.002 0.000 0.000
Residential Medium
1.204 25.489 17.027 1.526 1.189
Density
Transportation 0.041 0.455 0.888 0.500 1.083
Utilities 0.003 0.022 0.041 0.052 0.068
Four-class 2016 % developed imperviousness, NDCCI/NBI/BAEI SVM and iso cluster fusion
Land Use Urban 1 (sq km) Urban 2 Urban 3 Urban 4
Commercial and Services 0.560 1.694 4.271 8.374
Industrial 0.020 0.252 0.502 2.356
Institutional 0.330 1.162 1.645 1.347
Open Land 0.048 0.041 0.058 0.074
Recreational 0.669 0.446 0.330 0.376
Residential High Density 0.209 2.221 4.577 2.404
Residential Low Density 0.009 0.059 0.002 0.000
Residential Medium
1.204 25.489 17.027 2.715
Density
Transportation 0.041 0.455 0.888 1.583
Utilities 0.003 0.022 0.041 0.120
Remote Sens. 2020, 12, 2503 22 of 28

Having established mapping conventions with NBI, BAEI, and NDCCI, further consideration
was given to the potential for establishing an infill urban growth prediction implementing well-known
land change geosimulation modeling methodology based on the unexplored potential of utilizing
BAEI as a shadow index.
Remote Sens. 2020, 12, x FOR PEER REVIEW 23 of 29
3.4. Geosimulating High-Rise Infrastructure Development
3.4. Geosimulating High-Rise Infrastructure Development
A time-series of BAEI rasters without feature extraction thresholding, including data from Landsat
A time-series
5 for 2003 and 2008ofand
BAEI rasters8 without
Landsat for 2016,feature extraction
clipped thresholding,
to the Brickell includingcaptures
neighborhood, data from
the
Landsat 5 for 2003 and 2008 and Landsat 8 for 2016, clipped to the Brickell neighborhood,
formation of shadows (see Figure 11a–c). Additional clipping was performed to isolate an area of captures
the formation of shadows (see Figure 11a–c). Additional clipping was performed to isolate an area of
interest (see Figure 11d–f). By focalizing the rasters to a one-cell radius circle, the shadows were
interest (see Figure 11d–f). By focalizing the rasters to a one-cell radius circle, the shadows were
grouped into simple circular features in a new raster, and a threshold that set a good visible fit for
grouped into simple circular features in a new raster, and a threshold that set a good visible fit for
highlighting the shadows seen in the BAEI images was applied. Simple four-class land use maps,
highlighting the shadows seen in the BAEI images was applied. Simple four-class land use maps,
compatible with the LCM module, were created for the 2003 and 2008 data through the manual
compatible with the LCM module, were created for the 2003 and 2008 data through the manual
interpretation of high-resolution 2016 basemap imagery available through ArcGIS and geoprocessing
interpretation of high-resolution 2016 basemap imagery available through ArcGIS and geoprocessing
to include the shadows. Before initiating the geosimulation, polynomial transition trends were mapped
to include the shadows. Before initiating the geosimulation, polynomial transition trends were
within LCM during the module’s model fitting routine to determine a series of polynomial transition
mapped within LCM during the module’s model fitting routine to determine a series of polynomial
trends that could serve as effective predictors of the transition from “business” to “high-rise shadow”
transition trends that could serve as effective predictors of the transition from “business” to “high-
(see Figure 12).
rise shadow” (see Figure 12).

Figure11.
Figure 11. BAEI
BAEI (as
(as aa shadow
shadow index)
index) change
change detection
detection time-series
time-series ofof the
the Brickell
Brickell neighborhood
neighborhood
derived from Landsat data; (a) 5 TM 20 November 2003, (b) 17 November 2008, The
derived from Landsat data; (a) 5 TM 11/20/03, (b) 11/17/08, (c) 8 OLI 10/22/16. (c) 8 area of interest
OLI 22 October
overlaid with a focalized classification (transparent) based on thresholding; (d) 5 TM 11/20/03.
2016. The area of interest overlaid with a focalized classification (transparent) based on thresholding; (e)
11/17/08. (f) 8 OLI 10/22/16.
(d) 5 TM 20 November 2003. (e) 17 November 2008. (f) 8 OLI 22 October 2016.

The assumption made was that the geosimulation need not predict the 2016 focalization exactly
and that the prediction will, at least, closely outline the area where new high-rise infrastructure has
been constructed. LCM features an accuracy assessment tool that will accept a classification map as
input, though the results of this analysis were left to visual assessment. The output raster was clipped
to the area of interest and vectorized. Figure 13 displays the geosimulation output as two maps: the
predicted land use classification for the area of interest and a simplified vector of the output classified
Remote Sens. 2020, 12, x FOR PEER REVIEW 24 of 29

of skyscrapers based on the reasonable fit of the projected vector. This analysis serves as a basic
example of effective modeling that can be achieved through the proper model fitting of a
geosimulation that
Remote Sens. 2020, 12, performs based on the classification of urban land use according to transition
2503 23 of 28
potentials.

Figure 12. Classifications, created by manual and semi-automated methods, of the Brickell
Figure 12. Classifications, created by manual and semi-automated methods, of the Brickell neighborhood
neighborhood derived5 from
derived from Landsat TM dataLandsat 5 TM
used as LCMdata used(a)as20LCM
inputs; inputs;2003,
November (a) 11/20/03, (b) 11/17/08.
(b) 17 November 2008.
Polynomial transition trends created in LCM; (c) 2nd order, (d) 3rd order.
Polynomial transition trends created in LCM; (c) 2nd order, (d) 3rd order.

The assumption made was that the geosimulation need not predict the 2016 focalization exactly
and that the prediction will, at least, closely outline the area where new high-rise infrastructure has
been constructed. LCM features an accuracy assessment tool that will accept a classification map
as input, though the results of this analysis were left to visual assessment. The output raster was
clipped to the area of interest and vectorized. Figure 13 displays the geosimulation output as two
maps: the predicted land use classification for the area of interest and a simplified vector of the output
classified as shadows. The predicted vector, referencing the 2016 data, provides an excellent visual fit for
the shadows rendered with BAEI. It is proposed that this is a novel method for predicting the formation
of skyscrapers based on the reasonable fit of the projected vector. This analysis serves as a basic
example of effective modeling that can be achieved through the proper model fitting of a geosimulation
that performs based on the classification of urban land use according to transition potentials.
Remote Sens. 2020, 12, 2503 24 of 28
Remote Sens. 2020, 12, x FOR PEER REVIEW 25 of 29

Figure13.
Figure 13.Results
ResultsofofLCM
LCM2016
2016prediction
prediction for
for Brickell
Brickell neighborhood
neighborhood area
area of
of interest
interest based
based on
on 2003
2003
and 2008 data; (a) predicted 2016 area of interest land cover, (b) vector derived from 2016 prediction
and 2008 data; (a) predicted 2016 area of interest land cover, (b) vector derived from 2016 prediction
referencingthe
referencing theformation
formationofofshadows
shadowsininthe
thearea
areaof
ofinterest.
interest.

4.4.Discussion
Discussion
Jieli
Jielietetal.
al.(2010)
(2010)[14]
[14] developed
developed NBI to to enhance
enhancethe thescale
scaleofofthethebrightness
brightnessofofurban
urban features
features in in
a
a satellite
satellite image.
image. ValueValueranges
rangesfor forthis
this index,
index,calculated
calculatedfrom from20162016Landsat
Landsat88data,
data,are are found
foundto to be
be
descriptive
descriptiveofofdiscrete
discreteland
landuse usetypes,
types,according
accordingto tofeature
feature brightness.
brightness. Objects capturing
capturing the the shapes
shapes
ofofMiami’s
Miami’surban urbanfeature
featuretypes
typeswerewerevisualized,
visualized, and and an an object
object delineating isolated formation
delineating an isolated formation of of
industrialland
industrial landwas wassuccessfully
successfully extracted.
extracted. ItItisisproposed
proposedthat thatNBI
NBIis isa utility forfor
a utility feature
featureextraction
extraction of
ofurban
urbanland land use types
use typesthrough
through OBIA.
OBIA.Then, building
Then, uponupon
building the potential for identifying
the potential discretediscrete
for identifying urban
land land
urban use types,
use types,the the
possibility
possibilityof grading
of grading urbanization
urbanization in in
terms
termsofofcollective
collectivedevelopment
developmentwas was
investigated. Bouzekri et al. (2015) [12] developed BAEI to enhance the
investigated. Bouzekri et al. (2015) [12] developed BAEI to enhance the brightness of urban features inbrightness of urban features
a in a satellite
satellite image. image. Time-series
Time-series analyses
analyses of Miamiof Miami visualized
visualized withindex,
with this this index,
spanningspanning 1985–2016,
1985–2016, yields
a yields
view of a view
linearofincrease.
linear increase.
Also, the Also, the success
success of aimage
of a nine nine image time-series
time-series analysis
analysis comparing
comparing the two the
two indices for the truncated area was successful in identifying a negative
indices for the truncated area was successful in identifying a negative correlation between them. It is correlation between them.
It is proposed
proposed that BAEIthat BAEI
can becan be utilized
utilized during during
urbanurban surface
surface quality quality analyses
analyses to describe
to describe the upward
the upward scale
scale of development. Samsudin et al. (2016) [20] developed NDCCI
of development. Samsudin et al. (2016) [20] developed NDCCI to enhance the presence and condition to enhance the presence and
condition of construction material. Therefore, despite similar spectral
of construction material. Therefore, despite similar spectral configurations, it was suggested that this configurations, it was
suggested
index mightthat this index
be more usefulmight be more
for urban usefulthan
mapping for urban
NDVI,mapping
which is than NDVI,
designed which is the
to enhance designed
presence to
enhance the presence and condition of vegetation. A comparison
and condition of vegetation. A comparison between the two indices, derived from 2016 Landsat 8 between the two indices, derived
fromwith
data, 2016FGDLLandsat 8 data,
Miami land with
useFGDL MiamiNDCCI
data found land use data found
possesses lessNDCCI possesses
within-class less within-class
variance compared to
variance
NDVI. compared
A visual to NDVI.ofAiso
comparison visual
clustercomparison of isoof
classifications cluster classifications and
NDCCI/NBI/BAEI of NDCCI/NBI/BAEI
NDVI/NBI/BAEI
and NDVI/NBI/BAEI
composites showed thecomposites
NDCCI compositeshowed the NDCCIMiami
rendered compositewithrendered Miami between
less saturation with less features
saturation of
between features
various intensity. of various intensity.
Furthermore,ititwas
Furthermore, wasdiscernable
discernable that that urbanization
urbanization can can bebe linearly
linearly rendered
rendered according
according to to three
three
dominant feature types: open space, residential, and business/transport areas. Therefore, SVM was
dominant feature types: open space, residential, and business/transport areas. Therefore, SVM was
used to classify the NDCCI composite in three stages of intensity based on training samples derived
used to classify the NDCCI composite in three stages of intensity based on training samples decided
from a comparison between the composite and 2016 high-resolution satellite imagery. In addition to
from a comparison between the composite and 2016 high-resolution satellite imagery. In addition to
functioning as a standalone method, it serves as a palette for more detailed cartographic analyses to
functioning as a standalone method, it serves as a palette for more detailed cartographic analyses to
maximize the amount of information extracted from satellite data. It was possible to add more detail
maximize the amount of information extracted from satellite data. It was possible to add more detail
by fusing the SVM classification with NLCD 2016 percent developed imperviousness data. The area
Remote Sens. 2020, 12, 2503 25 of 28

by fusing the SVM classification with NLCD 2016 percent developed imperviousness data. The area of
Miami classified in NLCD 2016 as medium intensity was fused with the SVM classification in addition
to the fifth class of the NDCCI composite iso cluster classification to correct any residual errors. Because
of the simple workflow, it is within reason to assert the proposed methodology, when rigorously
applied, may facilitate the production of maps on any scale.
There is global consideration for the use of LULC maps, such as those created during this research,
as inputs into predictive geosimulations [1–3]. For example, Sun et al. (2007) [1] simulate the internal
growth of Calgary, Alberta, Canada utilizing discrete classes derived from city land use data. Object
maps derived from NBI may be implemented to facilitate this type of geostatistical modeling. In this
study, BAEI rasters without feature extraction thresholds were focalized to magnify the presence of
detected high-rise shadows in Miami’s Brickell neighborhood. Classifications derived from manual
interpretation and focal analysis of 2003 and 2008 Landsat 5 data were used as inputs for an MLP ANN
Markov Chain geosimulation, as well as polynomial transition trends derived from the classifications
for predictors. The geosimulation generated a land use prediction that closely circled the formation of
high-rise shadows in 2016 Landsat 8 data. The proposed methodology may be replicated to predict
the formation of any relevant discrete occurrence that may be otherwise difficult to predict, given the
isolated nature of the occurrence. Additional model fitting may be useful in refining the prediction,
such as the implementation of other predictor gradients.
Challenges for replicating the study methods include the availability of GIS and remote
sensing data, evaluating OBIA settings, managing the presence of vegetation, fitting an iso cluster
classification to suitably capture the intensity of features, establishment of correct SVM training
samples, and establishing a geosimulation area of interest. While the static NBI, BAEI, composite iso
cluster, and composite SVM maps can be generated globally, it is understood that the fusion maps can
only be generated if a percent developed imperviousness product exists for the area being mapped.
For the study area of Miami, the spectral response among built-up indices was found to be sensitive to
differences between land use types and to changes that occurred over a time-series. It must be duly
noted that the study area was found to be comprised of three primary types: open spaces, residential,
and commercial development. The observable sensitivity may be attributed to the composition of the
city’s morphology as well as lighting conditions attributed to the city’s position and the time of data
acquisition. Henceforward future work must consider the sensitivity of spectral responses in urban
areas with morphological compositions that differ from Miami, as well as the influence of differing
lighting conditions.

5. Conclusions
Rapid urbanization stimulates a high demand for static mapping and geosimulations to support
legislative and planning purposes. A caveat exists where LULC maps should be synthesized through a
process of maximizing the amount of information that can be acquired from a satellite image. While this
is not the first study to research the potential for incorporating spectral enhancements in classification
processes delineating multiple states of urbanization, the paper systematically summarizes specific
approaches to various aspects of urban analysis. It discusses the advantages and limitations of existing
indices. The power of remote sensing and GIS were coupled to conceptualize new urban land use
classification capabilities utilizing the NBI, BAEI, and NDCCI spectral enhancements. LULC maps
generated without the details that may be acquired from these enhancements may be less useful
in application.
Today, the trend of global urbanization has negatively impacted Earth by triggering climate
change, lowering water quality, and reducing natural landscape, among other problems. In response,
there is an increased need for frequently updated LULC maps, and smart growth policies have been
implemented by those responsible for managing the development of cityscapes. The principle of infill
development, the redevelopment of land within a cityscape, will become more important because of
the goal of smart growth policies to build compact, high population urban areas instead of sprawling
Remote Sens. 2020, 12, 2503 26 of 28

outward. When combined with the analytical capabilities of GIS, remote sensing data may be refined to
focus on a specific geography. Miami, a city that exists in a sensitive coastal environment, continues to
grow at a rapid pace. We should consider using the city as a study area for research on urban geographic
information science. Maps derived from this research may be advantageous for strengthening all
purposes that rely on LULC maps.
Significant to this study, NBI was demonstrated to possess the unique property of identifying
discrete urban land use types as segmented objects. Urban LULC maps may be created, through the
process of OBIA, based on the geometric configuration of discrete land use types rendered with NBI.
BAEI was successfully implemented to visualize and scale the growth of an entire city as well as an
area truncated of urban canopy. Such a rendering of urban surface quality may be useful for important
tasks such as predicting stages of development bordering natural areas, where intense development
may be detrimental to the environment. NDCCI, when composited with NBI and BAEI, can be utilized
to classify stages of urban intensity effectively.
The development of urban growth geosimulation models meant to bolster management efforts
may become the research frontier for all efforts to mitigate human-driven environmental impacts.
Satellite data are generated continuously and serve as an excellent basis for the LULC maps necessary
for geosimulations to operate. The urban land use classification methods introduced in this paper
have the potential to serve as inputs into urban growth geosimulations. NBI can be used to extract
predictable geometries, BAEI can be reclassified into value ranges describing predictable gradient
development, and the NDCCI/NBI/BAEI composite maps can be used to facilitate predictions of
fluctuating urban intensity. While it is beyond the scope of this paper to give further consideration to
these ideas, the novel method incorporating the BAEI as a shadow index is a clear example of the type
of geostatistical model that can be created to monitor a city’s internal growth. Forward, it should also
serve as a primary example of the type of success that may be gleaned from focalized geosimulation
models in general.

Author Contributions: P.L. developed methodology, performed formal analysis, validation, visualization,
interpreted results and wrote the paper; L.B. and E.H. aided to methodological design and contributed to final
manuscript. All authors have read and agreed to the published version of the manuscript.
Funding: This research received no external funding.
Acknowledgments: The authors thank San Francisco State University Department of Geography & Environment
for time allowed to complete this manuscript. We thank all reviewers for their comments towards improving
this manuscript.
Conflicts of Interest: The authors declare no conflict of interest.

References
1. Sun, H.; Forsythe, W.; Waters, N. Modeling Urban Land Use Change and Urban Sprawl: Calgary, Alberta,
Canada. Netw. Spat. Econ. 2007, 7, 353–376. [CrossRef]
2. Guan, D.; Li, H.; Inohae, T.; Su, W.; Nagaie, T.; Hokao, K. Modeling urban land use change by the integration
of cellular automaton and Markov model. Eco. Mod. 2011, 222, 3761–3772. [CrossRef]
3. Rahimi, A.A. A methodological approach to urban land-use change modeling using infill development
pattern—A case study in Tabriz, Iran. Ecol. Process. 2016, 5, 1–15. [CrossRef]
4. Seto, K.C.; Güneralp, B.; Hutyra, L.R. Global forecasts of urban expansion to 2030 and direct impacts on
biodiversity and carbon pools. Proc. Nat. Acad. Sci. USA 2012, 109, 16083–16088. [CrossRef]
5. Mohan, M.; Kandya, A. Impact of Urbanization and Land-use/Land-cover Change on Diurnal Temperature
Range: A Case Study of Tropical Urban Airshed of India Using Remote Sensing Data. Sci. Total Environ.
2015, 506–507, 453–465. [CrossRef] [PubMed]
6. Nelson, K.C.; Palmer, M.A.; Pizzuto, J.E.; Moglen, G.E.; Angermeier, P.L.; Hilderbrand, R.H.; Dettinger, M.;
Hayhoe, K. Forecasting the combined effects of urbanization and climate change on stream ecosystems:
From impacts to management options. J. App. Ecol. 2009, 46, 154–163. [CrossRef] [PubMed]
Remote Sens. 2020, 12, 2503 27 of 28

7. Zia, S.; Shirazi, S.A.; Bhalli, M.N.; Kausar, S. The impact of urbanization on mean annual temperature of
lahore metropolitan area, pakistan. Pak. J. Sci. 2015, 67, 301–307.
8. Han, G.; Xu, J. Land Surface Phenology and Land Surface Temperature Changes Along an Urban-Rural
Gradient in Yangtze River Delta, China. Environ. Manag. 2013, 52, 234–249. [CrossRef] [PubMed]
9. Pei, F.; Wu, C.; Liu, X.; Li, X.; Yang, K.; Zhou, Y.; Wang, K.; Xu, L.; Xia, G. Monitoring the vegetation activity
in China using vegetation health indices. Agric. For. Meteorol. 2018, 248, 215–227. [CrossRef]
10. Tang, H.; Li, Z.; Zhu, Z.; Chen, B.; Zhang, B.; Xin, X. Variability and Climate Change Trend in Vegetation
Phenology of Recent Decades in the Greater Khingan Mountain Area, Northeastern China. Remote Sens.
2015, 7, 11914–11932. [CrossRef]
11. As-syakur, A.; Adnyana, I.; Arthana, I.; Nuarsa, I. Enhanced Built-up and Bareness Index (EBBI) for Mapping
Built-up and Bare Land in an Urban Area. Remote Sens. 2012, 4, 2957–2970. [CrossRef]
12. Bouzekri, S.; Lasbet, A.; Lachehab, A. A New Spectral Index for Extraction of Built-up Area Using Landsat-8
Data. J. Indian Soc. Remote Sens. 2015, 43, 867–873. [CrossRef]
13. He, C.; Shi, P.; Xie, D.; Zhao, Y. Improving the normalized difference built-up index to map urban built-up
areas using a semiautomatic segmentation approach. Remote Sens. Lett. 2010, 1, 213–221. [CrossRef]
14. Jieli, C.; Manchun, L.; Yongxue, L.; Chenglei, S.; Wei, H. Extract residential areas automatically by New
Built-up Index. In Proceedings of the 2010 18th International Conference on Geoinformatics, Beijing, China,
18–20 June 2010; Volume 1.
15. Rasul, A.; Balzter, H.; Ibrahim, G.R.F.; Hameed, H.M.; Wheeler, J.; Adamu, B.; Ibrahim, S.; Najmaddin, P.M.
Applying Built-up and Bare-Soil Indices from Landsat 8 to Cities in Dry Climates. Land 2018, 7, 81. [CrossRef]
16. Waqar, M.M.; Mirza, J.F.; Mumtaz, R.; Hussain, E. Development of new indices for extraction of built-up area
and bare soil from landsat. Open Access Sci. Rep. 2012, 1, 1–4.
17. Xu, H. A new index for delineating built-up land features in satellite imagery. Int. J. Remote Sens. 2008, 29,
4269–4276. [CrossRef]
18. Zha, Y.; Gao, J.; Ni, S. Use of normalized difference built-up index in automatically mapping urban areas
from TM imagery. Int. J. Remote Sens. 2003, 24, 583–594. [CrossRef]
19. Benkouider, F.; Abdellaoui, A.; Hamami, L. New and Improved Built-up Index Using SPOT Imagery:
Application to an Arid Zone (Laghouat and M’Sila, Algeria). J. Indian Soc. Remote Sens. 2019, 47, 185–192.
[CrossRef]
20. Samsudin, S.H.; Shafri, H.Z.M.; Hamedianfar, A. Development of spectral indices for roofing material
condition status detection using field spectroscopy and WorldView-3 data. J. Appl. Remote Sens. 2016, 10,
025021. [CrossRef]
21. Gu, L.; Cao, Q.; Ren, R. Building extraction method based on the spectral index for high-resolution remote
sensing images over urban areas. J. Appl. Remote Sens. 2018, 12, 1. [CrossRef]
22. Peeroo, U.; Idrees, M.; Saeidi, V. Building extraction for 3D city modelling using airborne laser scanning data
and high-resolution aerial photo. S. Afr. J. Geomat. 2017, 6, 363–376. [CrossRef]
23. United Nations Department of Economic and Social Affairs (UN DESA): World Population Projected to
Reach 9.8 Billion in 2050, and 11.2 Billion in 2100. Available online: https://www.un.org/development/desa/
en/news/population/world-population-prospects-2017.html (accessed on 15 December 2019).
24. Abbas, A.W.; Ahmad, N.; Abid, S.A.R.; Khan, M.A.A. K-means and ISODATA Clustering Algorithms for
Landcover Classification Using Remote Sensing. Sindh Univ. Res. J. SURJ (Sci. Ser.) 2016, 48, 315–318.
25. Igun, E. Analysis and Sustainable Management of Urban Growth’s Impact on Land Surface Temperature in
Lagos, Nigeria. J. Remote Sens. GIS 2017, 6, 212. [CrossRef]
26. Pal, M.; Mather, P. Support vector machines for classification in remote sensing. Int. J. Remote Sens. 2005, 26,
1007–1011. [CrossRef]
27. Ettehadi, O.E.; Kaya, S.; Elif, S.; Alganci, U. Separating Built-up Areas from Bare Land in Mediterranean
Cities Using Sentinel-2A Imagery. Remote Sens. 2019, 11, 345. [CrossRef]
28. Cabral, P. Délimitation d‘aires urbaines à partir d‘une image Landsat ETM+: Comparaison de méthodes de
classification. Can. J. Remote Sens. 2007, 33, 422–430. [CrossRef]
29. Beck, L.; Hutchinson, R.; Zauderer, C. A comparison of greenness measures in two semi-arid grasslands.
Clim. Chang. 1990, 17, 287–303. [CrossRef]
Remote Sens. 2020, 12, 2503 28 of 28

30. Jin, S.; Homer, C.; Yang, L.; Danielson, P.; Dewitz, J.; Li, C.; Zhu, Z.; Xian, G.; Howard, D. Overall Methodology
Design for the United States National Land Cover Database 2016 Products. Remote Sens. 2019, 11, 2971.
[CrossRef]
31. Environmental Systems Research. ArcGIS Pro: Version 2.5.1; Environmental Systems Research Institute
(ESRI): Redlands, CA, USA, 2020.
32. Hexagon Geospatial. ERDAS Imagine: Version 16.5.0; Hexagon Geospatial: Madison, AL, USA, 2018.
33. Clark Labs, Clark University. TerrSet Geospatial Monitoring and Modeling Software: Version 18; Clark Labs,
Clark University: Worcester, MA, USA, 2015.
34. World Population Review: Miami, Florida Population 2020. Available online: http://worldpopulationreview.
com/us-cities/miami-population/ (accessed on 12 April 2020).
35. Kim, J.; Steiner, R.; Yang, Y. The Evolution of Transportation Concurrency and Urban Development Pattern
in Miami-Dade County, Florida. Urban Aff. Rev. 2014, 50, 672–701. [CrossRef]
36. United States Geological Survey (USGS): Landsat Surface Reflectance. Available online: https://www.usgs.
gov/land-resources/nli/landsat/landsat-surface-reflectance (accessed on 12 April 2020).
37. Phiri, D.; Morgenroth, J. Developments in Landsat Land Cover Classification Methods: A Review. Remote Sens.
2017, 9, 967. [CrossRef]
38. Chrysoulakis, N. Estimation of the all-wave urban surface radiation balance by use of ASTER multispectral
imagery and in situ spatial data. J. Geophys. Res. Atmos. 2003, 108. [CrossRef]
39. Dixit, A.; Goswami, A.; Jain, S. Development and Evaluation of a New “Snow Water Index (SWI)” for
Accurate Snow Cover Delineation. Remote Sens. 2019, 11, 2774. [CrossRef]
40. Hochmair, H.H.; Gann, D.; Benjamin, A.; Fu, J. Miami-Dade County Urban Tree Canopy Assessment; GIS Center,
Florida International University: Miami, FL, USA, 2016.
41. Ball, G.H.; Hall, D.J. Isodata: A Method of Data Analysis and Pattern Classification; Stanford University:
Menlo Park, CA, USA, 1965.
42. Cortes, C.; Vapnik, V. Support-vector Networks. Mach. Learn. 1995, 20, 273–297. [CrossRef]
43. Ustuner, M.; Sanli, F.B.; Abdikan, S. Balanced vs imbalanced training data: Classifying rapideye data with
support vector machines. Int. Arch. Photogramm. Remote Sens. Spat. Inf. Sci. 2016, XLI-B7, 379–384.
[CrossRef]
44. Chen, D.; Stow, D. The Effect of Training Strategies on Supervised Classification at Different Spatial
Resolutions. Photogramm. Eng. Remote Sens. 2002, 68, 1155–1161.
45. Marius, P.; Balas, V.; Perescu-Popescu, L.; Mastorakis, N. Multilayer Perceptron and Neural Networks.
WSEAS Trans. Circuits Syst. 2009, 8, 579–588.

© 2020 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access
article distributed under the terms and conditions of the Creative Commons Attribution
(CC BY) license (http://creativecommons.org/licenses/by/4.0/).