Download as pdf or txt
Download as pdf or txt
You are on page 1of 17

Journal of the Indian Society of Remote Sensing (July 2022) 50(7):1257–1273

https://doi.org/10.1007/s12524-022-01526-7(0123456789().,-volV)(0123456789().
,- volV)

RESEARCH ARTICLE

A Comparison of Discrete and Continuous Metrics for Measuring


Landscape Changes
Ehsan Rahimi1 • Shahindokht Barghjelveh2 • Pinliang Dong3

Received: 28 December 2020 / Accepted: 19 February 2022 / Published online: 13 March 2022
Ó Indian Society of Remote Sensing 2022

Abstract
Quantification of landscape patterns in a correct way is an important issue in landscape ecology, which has recently
attracted much attention from landscape ecologists. It is believed that Patch-Corridor-Matrix or discrete model does not
take into account the continuous heterogeneity and has limitations and problems that undermine the validity of the results.
Various continuous methods have been developed to overcome the problems associated with the discrete approach for
measuring landscape fragmentation. Continuous methods use remotely sensed data directly for quantifying landscape
pattern changes. In this regard, the purpose of this study is to compare landscape metrics obtained from the discrete model
and alternative continuous metrics, including spatial autocorrelation indices, Fourier transforms, and surface metrics. To
achieve this goal, we used two subsets that were different in terms of urban and agricultural changes. We measured
temporal changes of three subsets between the years 2013 and 2020. Results showed that percentage of landscape
(PLAND) as a landscape metric had a good statistical relationship with Getis of the normalized difference vegetation index
(NDVI) (R2 = 95.9%) and a moderate relationship with Getis of the built-up index (PNBI) (R2 = 43.7%) as alternative
metrics. The results of discrete metrics showed that in 2013, the spatial patterns of urban and cultivated farms patches were
more fragmented than in 2020. Also, the area of cultivated farms increased by 3 to 6 percent and urban areas between 1 and
8 percent in 2020. All continuous metrics showed that the subsets in the study changed in 2020. However, Fourier
transforms could not determine the magnitude of these changes. Our results also showed that landscape metrics have some
drawbacks for measuring landscape patterns that some of them can be resolved by continuous metrics. Generally, the use of
discrete and continuous metrics depends on factors like the scale of the study, time and budget available, desired ecological
processes, and the degree of heterogeneity of the landscape of interest.

Keywords Landscape metrics  Continuous metrics  Landscape changes  Landscape ecology

Introduction

The estimation of spatial heterogeneity in a correct way


& Ehsan Rahimi and scale according to the species and process under study
ehsanrahimi666@gmail.com
is an important challenge in studying and managing land-
Shahindokht Barghjelveh scapes (Fan & Myint, 2014). Success in estimating rela-
shjelveh@gmail.com
tionships between patterns and processes is dependent on
Pinliang Dong the correct identification of landscape heterogeneity in the
Pinliang.Dong@unt.edu
manner that is relating to the organism or the process. With
1
Environmental Sciences Research Institute, Shahid Beheshti this view, landscape ecologists apply a discrete approach
University, Tehran, Iran that considers landscapes as a collection of discrete patches
2
Associate Professor of Environmental Sciences, (Forman, 2014). The patch-corridor-matrix (PCM) model
Environmental Sciences Research Institute, Shahid Beheshti was developed in the 1980s to describe landscape structure
University, Tehran, Iran as a collection of homogenous areas that have been mapped
3
Department of Geography and the Environment, University discretely. The quantification of landscape patterns based
of North Texas, Denton, TX, USA

123
1258 Journal of the Indian Society of Remote Sensing (July 2022) 50(7):1257–1273

on this model requires a discrete assumption of the land- ability to describe the relationship between patterns and
scape with a suitable thematic resolution (the number of processes. Mondal (2011) also used surface metrics to
discrete classes on the classification map) (McGarigal & quantify continuous landscape heterogeneity and showed
Cushman, 2005). that surface metrics were capable of describing the spatial
Today, many discrete metrics (known as landscape and non-spatial features of the landscape. Many surface
metrics) have been developed to quantify landscape com- metrics show consistent scaling relations according to
position and configuration based on the PCM model, but changing resolution and several can accurately predict
most of them are unable to characterize ecological pro- values at finer resolutions (Frazier, 2016). Recently,
cesses (Frazier & Kedron, 2017; Kedron et al., 2019) and Kedron et al. (2018) have investigated the performance and
there are several other limitations related to this model applicability of surface metrics in describing different
(Gustafson, 2019; Hagen-Zanker, 2016; Park & Guldmann, forest landscapes and found that not all surface metrics are
2020). One of the problems is that the errors related to the suitable for landscape analysis.
land-use classification are inevitable and may lead to unreal Fourier transforms have also been introduced as a tool
results (Shao & Wu, 2008). The constraints associated with for describing changes in landscape patterns over time. The
the classification of continuous variables of the landscapes use of this tool in landscape studies was first undertaken by
into discrete classes have reduced the functionality and Rocchini, Metz, et al. (2013) to investigate landscape
validity of these metrics (Frazier & Kedron, 2017; Kedron fragmentation using satellite imagery in different periods.
et al., 2018; Park & Guldmann, 2020). Therefore, the The results of their study showed that Fourier analysis was
uncertainty associated with classification can compromise well able to detect landscape fragmentation in different
the reliability of landscape metrics derived from the the- periods, but they did not provide any information on how
matic maps (Cockx et al., 2014; Fan & Myint, 2014; much these changes were and where these changes
Kedron et al., 2018; Rocchini, Foody, et al., 2013). Several occurred in the landscape. Other alternative metrics are the
studies also showed that the accuracy of landscape metrics local spatial autocorrelation indices. These indices have
is strongly related to the accuracy of classified maps (Li & been used by Fan and Myint (2014) to quantify landscape
Wu, 2004; Shao & Wu, 2008). Many other factors affect patterns, and their performance has been compared with
the accuracy and applicability of landscape metrics, like landscape metrics in the quantification of urban landscape
data source accuracy, scale effects, and ecological inter- fragmentation. The results of the comparison showed the
pretation (Frazier & Kedron, 2017; Liu et al., 2013). efficiency of these indices in the quantification of land-
By raising awareness of the problems and limitations of scape heterogeneity.
the PCM model, alternative metrics to landscape structure The use of patterns and metrics to evaluate landscapes in
quantification are needed (Frazier & Kedron, 2017; Kedron a continuous framework is rare, and so far few comparisons
et al., 2018, 2019). McGarigal and Cushman (2005) have been made between the use of continuous patterns and
introduced the gradient model as an alternative approach to the discrete pattern of the PCM model (Fan & Myint, 2014;
landscape structure quantification. Instead of mapping Kedron et al., 2018). Various studies that explore the
homogeneous and discrete areas, the gradient model rep- landscape patterns in the continuous framework have all
resents landscape structure based on continuous data in mentioned the shortcomings of discrete models and intro-
which the only discrete units are pixels of a raster. In this duced alternative metrics (Frazier, 2016; Kedron et al.,
model, heterogeneity is shown as a three-dimensional (3D) 2018; Lausch et al., 2015; McGarigal et al., 2009; Rahimi
surface that can represent any desired ecological feature et al., 2021). However, it is often not possible to compare
(McGarigal & Cushman, 2005). Continuous representation the results of continuous and discrete approaches directly
of a landscape contains much more information and because the output of some methods is an image and others
therefore presents a realistic representation of landscapes a number. For example, the output of landscape and surface
(Lausch et al., 2015). metrics is a number, but the output of local spatial auto-
Nowadays, in landscape ecological studies, much correlation indices and Fourier transform is an image.
attention has been paid to continuous metrics to measure Moreover, the ability of these methods to detect small
landscape fragmentation. Various studies have investigated changes has not been well investigated. Therefore, in this
the landscapes in a continuous framework to overcome the study, we are interested in comparing these methods for
limitations of the PCM model. One set of metrics that are change detection. The objectives of this paper are to
used as alternatives for PCM-based landscape metrics in address the following questions:
quantifying landscape patterns is surface metrics. These
(1) What are the weaknesses of the methods used as
metrics were used by (McGarigal et al., 2009) in a study to
alternatives to the PCM model?
describe landscape patterns. They acknowledged that the
surface metrics have unique characteristics and a high

123
Journal of the Indian Society of Remote Sensing (July 2022) 50(7):1257–1273 1259

(2) What are the strengths of these alternative methods 2. These subsets involve small changes in the built-up areas
for quantifying landscape patterns? and agricultural lands. Therefore, the efficiency of alter-
(3) Have there been any suitable alternatives to over- native metrics in detecting small changes can be well
come the drawbacks of the PCM model? determined.
The data used in this study consist of two satellite
images acquired from Landsat 8 OLI satellite (Table 1).
Landsat 8 OLI images have a spatial resolution of 30 m
Study Area
and a 12-bit radiometric precision. The first image was
taken on 11/6/2013 and the second image on 15/6/2020,
Figure 1 shows the location of the study area in Tehran
and two subsets were selected from each of these images.
province of Iran. In this figure, two subsets are separated to
Images were acquired during the vegetation growing sea-
determine the effectiveness of Fourier transforms, surface
sons and are freely available from the Landsat archive from
metrics, and local spatial autocorrelation indices in com-
the United States Geological Survey (USGS) (http://glovis.
parison with the landscape metrics to quantify landscape
usgs.gov). All data were used for the classification of the
changes. The area of subset 1 (Fig. 1a) is 100 hectares and
subsets and extraction of vegetation indices in a change
subset 2 (Fig. 1b) is 220 hectares that are located in the
detection manner over seven years (2013–2020) to deter-
west of Tehran province. Subset 1 is different from subset 2
mine the efficiency of all continuous and discrete methods
in terms of the area of urban and cultivated farms.
for measuring small changes in the subsets.
Approximately, 16% and 18% of subsets 1 are covered by
agriculture and urban covers, respectively, while cultivated
farms constitute 34% and urban areas cover 16% of subset

Fig.1 Location of study areas in Tehran province of Iran in the year 2020. a subset 1 and b subset 2. The green areas show the cultivated farms,
and the white areas show the urban features

123
1260 Journal of the Indian Society of Remote Sensing (July 2022) 50(7):1257–1273

Table 1 Details of satellite data used in this study as input, they can examine covers that can be distinguished
Date Path Row Satellite Sensor Resolution
by these indices. In this study, only two urban and agri-
culture cover types were identifiable by vegetation indices,
11/6/2013 164 35 Landsat 8 OLI 30 m and therefore, these two cover types were considered in the
15/6/2020 164 35 Landsat 8 OLI 30 m present study. The urban class included built-up areas, and
impervious surfaces like roads, parking areas, and cement
surfaces. The agriculture class consists only of cultivated
farms, which can be seen in Fig. 1 in green. Barren lands,
Methods uncultivated, and burned areas were all classified in the
other category. Five hundred random points were generated
Figure 2 shows the diagram comparing discrete and con- for each subset to obtain overall accuracy and kappa
tinuous metrics. This figure shows that landscape metrics coefficients of the classified maps. Reference points were
are used in discrete methods. The input of landscape mainly acquired from Google Earth images.
metrics is classified land-use maps that in the present study,
these maps were prepared using Landsat images. On the Measuring Landscape Changes
other hand, there are continuous methods that include
spatial autocorrelation indices, surface metrics, and Fourier Measuring landscape changes in this study is based on four
transforms. Unlike landscape metrics, continuous metrics methods including landscape metrics, local spatial auto-
inputs are mainly vegetation indices such as NDVI and correlation indices, surface metrics, and Fourier trans-
PNBI for surface metrics and spatial autocorrelation indi- forms. Since the output types of discrete and continuous
ces. The first principal component (PCA1) derived from the methods are different from each other, it is difficult to
images was used as input for Fourier transforms to inves- make a direct comparison between landscape metrics and
tigate changes in landscape patterns. The efficiency of alternative methods. For example, the outputs of landscape
discrete and continuous metrics was compared for change and surface metrics are numbers that cannot be compared
detection between 2013 and 2020. The methodology with output images of local spatial autocorrelation indices
flowchart is shown in Fig. 2. and Fourier transforms. To solve this problem, we tested
the ability of these methods to detect temporal variations in
Image Processing two separate subsets over seven years. We consciously
chose landscapes that had little changes in urban and
We used four classified maps for calculating landscape agricultural areas during these seven years. We measured
metrics and comparing these metrics with alternative the landscape changes by all discrete and continuous
methods. Using the maximum likelihood algorithm, methods and compared the results of these methods with
Landsat images of two subsets in the years 2013 and 2020 one another. We used classified maps to calculate changes
were classified into two classes including built-up and non- by landscape metrics and used vegetation indices as input
built-up. Since continuous methods use vegetation indices to continuous methods to detect the temporal changes of
urban and agriculture covers.

Fig. 2 Methodology
flowchart of comparing discrete Comparison of discrete and
and continuous metrics for continuous metrics
measuring landscape changes
Discrete metrics Continuous metrics

Landscape metrics Autocorrelation indices


Input
Surface metrics
Land use/cover

Fourier transforms
Change detection Input

NDVI, PNBI, PCA1

Statistical analysis Change detection

123
Journal of the Indian Society of Remote Sensing (July 2022) 50(7):1257–1273 1261

Landscape Metrics (Band 4) (Fan & Myint, 2014). This index is the most
widely used index for many different applications, ranging
To calculate land-use changes in seven years, we used six from vegetation monitoring to urban sprawl (Nolè et al.,
landscape metrics (Table 2) including the Percentage of 2014). One of the indices used for the study of urban
Landscape (PLAND), Large Patch Index (LPI), Number of phenomena is the built-up index (PNBI). The built-up
Patches (NP), Patch Density (PD), Effective Mesh Size index is a useful index to identify impervious surfaces like
(MESH), Aggregation Index (AI), Splitting Index (SPLIT) asphalt and concrete (Nolè et al., 2014). Values generated
available in Fragstats 4.2 software (McGarigal et al., 2002). using the NDVI and PNBI indexes range from -1 to 1. The
To compare local spatial autocorrelation indices and formulas for calculating NDVI and PNBI are as follows:
landscape metrics, it is necessary to visualize the output of NIR  RED
landscape metrics using the moving window technique. NDVI ¼ ð1Þ
NIR þ RED
Some studies have suggested that a window size of five
PCA1  NIR
cells is the best fit for measuring spatial patterns in a PNBI ¼ ð2Þ
PCA1 þ NIR
landscape by moving window method (Buyantuyev et al.,
2010; Fan & Myint, 2014), and therefore, a window size of where Red and NIR stand for the spectral reflectance
5 by 5 pixels was used for calculating landscape metrics. measurements acquired in the red (visible) and near-in-
The output of the moving window technique is a raster in frared regions, respectively, and PCA1 stands for the first
which each pixel represents the calculated value of the principal component of all bands.
landscape metrics. These landscape metrics were calcu-
lated at the class level for agriculture and built-up areas in Local Spatial Autocorrelation Indices
the subsets.
Spatial autocorrelation is an indicator that measures a
Vegetation Indices phenomenon in which closely related observations are
more similar than those that are farther away (Bolliger
For measuring landscape patterns in the continuous et al., 2007). Several indices have been developed to
framework, we need to use continuous indices that reflect measure the spatial autocorrelation to determine the exis-
landscape features. In this regard, we used two indexes tence and severity of clusters in the distribution. The global
including the Normalized Difference Vegetation Index spatial autocorrelation indices measure the total amount of
(NDVI) and the built-up index (PNBI) as alternative indi- data correlation across the study area, but local indicators
cators of landscape features. The normalized difference of spatial autocorrelation allow us to locate clustered pix-
vegetation index is effective for identifying green vegeta- els, by measuring how many features inside the fixed
tion biomass due to strong absorption in the red region neighborhood are homogeneous (Anselin, 1995). In this
(Band 3) and a strong reflection in the near-infrared band study, we used Getis–Ord (Getis & Ord, 1992) that

Table 2 Descriptions of the selected landscape metrics


Metric Equation Range Explanation
P
PLAND aij 0 B PLAND B 100 Shows the percentage of landscape occupied by a cover
A ð100Þ
maxaij
LPI A ð100Þ 0 B LPI B 100 Shows the percentage of the largest patch of a cover
P  1 
MESH A2IJ The ratio of cell size to landscape area B MESH B total High values of MESH indicate a low degree of
A 10000 landscape area (A) landscape fragmentation
 
AI gii
ð100Þ 0 B AI B 100 High values of AI indicate a low degree of landscape
max!gii
fragmentation
SPLIT PA2 2 1 B SPLIT B number of cells in the landscape area High values of SPLIT indicate a high degree of
AIJ
squared landscape fragmentation
ni
PD A ð10000Þð100Þ PD [ 0, constrained by cell size High values of PD indicate a high degree of landscape
fragmentation
aij = area (m2) of patch; A = total landscape area (m2); ni = number of class i patches in the landscape; eij = total length (m) of edges of patch ij,
including landscape boundary; aij c = area (m2) within patch ij separated from its boundary by a user-specified buffer width (m); gii = the
number of adjacencies (contiguity) between pixels of patch class i; max gii = maximum possible number of adjacencies among pixels of patches
of class i (McGarigal et al., 2002)

123
1262 Journal of the Indian Society of Remote Sensing (July 2022) 50(7):1257–1273

compares pixel values at a given location with those pixels to quantify landscape patterns of PNBI and NDVI indices
at a lag, d, from the original pixel at location i. The between years 2013 and 2020 (Table 3).
function by Getis and Ord (1992) is represented by the Surface skewness (Ssk), surface kurtosis (Sku), and
following equation: valley fluid retention index (Svi) can be interpreted as
GetisðdÞ measures of the dominance of landscape patterns, similar to
Pn Pn dominance metrics in the PCM model. Ssk shows the
i¼1 wi ðdÞxi  xi i¼1 wi ðdÞ asymmetry of the surface height distribution. A surface
¼ rffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
h P Pn 2 i
ðN  1Þ ni¼1 wi ðdÞ  with an equal number of peaks and valleys has zero
SðiÞ i¼1 wi ðdÞ =ðN  2Þ
skewness and a surface with the domination of high values
ð3Þ (peaks) with occasional valleys has a negative skewness,
while a surface with the domination of low values (valleys)
where N is the total pixel number, i stands for pixel i, Xi
with occasional peaks has positive skewness. Conse-
and Xj are intensities in points i and j (with i = j), value,
quently, the amount of skewness depends on the position-
wij (d) represents a weight matrix which varies according to
P ing of the majority of the values above the surface
distance, S1i= Nj¼1 wij 2 .
(negative skewness) or below the surface (positive skew-
The high value of Getis means a positive correlation for ness) relative to the average height.
high values of intensity, while a low value of the index A surface with a relatively even distribution of heights
means a positive correlation for low values of intensity above and below the mean has low kurtosis (Sku \ 3), and
(Lanorte et al., 2013). We used ENVI5.3 software for the a surface with relatively few areas above or below the
calculating local spatial autocorrelation index. To be con- mean has high kurtosis (Sku [ 3). Therefore, high kurtosis
sistent with the moving window analysis conducted above, indicates a landscape with a dominant surface height, akin
we used a 5 9 5 window size for calculating Getis statis- to a ‘matrix’ under the PCM model of landscape structure.
tics. The Getis statistics was applied to vegetation and The surface kurtosis (Sku) metric only shows the domi-
built-up indices derived from Landsat data for the years nance of the landscape and does not provide information on
2013 and 2020. A sample of 2500 randomly generated which cover is dominant. Therefore, this metric should be
points was utilized to evaluate the relationship between interpreted with Ssk and Svi metrics. The average rough-
landscape metrics and Getis statistics at the class level. To ness (Sa) metric measures the average absolute deviation of
detect changes between different years using Getis statis- the surface heights from the mean. This metric can be
tics, we applied Getis statistics to NDVI and PNBI indices interpreted as a non-spatial metric of landscape diversity,
and compared them with the moving window images of analogous to the PCM-based diversity metrics. Larger
agriculture and urban covers, respectively. For change values represent an increasing range of values in the sur-
detection, we calculated the probability density functions face attribute (akin to increasing patch richness). Valley
of these images. fluid retention index (Svi), like Ssk and Sku, is also a
metric of the shape of the surface height profile.
Surface Metrics
Fourier Transforms
Similar to the landscape metrics, the landscape gradients
also have spatial and non-spatial components. Spatial The Fourier transforms are well known in mathematics as
components refer to vertical and horizontal variations, but linear transformations from the time domain to the fre-
non-spatial components only refer to vertical variability in quency domain. If we assumed that f (x) be a continuous
the surface (Mondal, 2011). Generally, non-spatial com- function described in a spatial domain. Based on the
ponents do not provide any information about the Fourier theorem (Fourier, 1822), every f (x) can be trans-
arrangement, location, or distribution of valleys and peaks formed into a continuum of sinusoidal functions of varying
on the surface. These metrics are similar to diversity frequency, as:
metrics in the PCM model and measure composition
Z
þ1
aspects of landscapes, not configuration (Kedron et al.,
2018; McGarigal et al., 2009). In contrast to the suite of F ðx Þ ¼ f ðxÞe2pixx dx ð4Þ
available patch metrics, the available surface metrics are all 1

structural. We applied some of these metrics calculated by where x = frequency, also known as radian frequency
software SPIP (Metrology, 2002) to examine their ability in since it is expressed in radians per spatial unit.
quantifying surface patterns. Four surface metrics includ- Extending Eq. 4 to two dimensions means the consid-
ing surface skewness (Ssk), surface kurtosis (Sku), average eration of a two-dimensional function such as a raster
roughness (Sa), valley fluid retention index (Svi) were used

123
Journal of the Indian Society of Remote Sensing (July 2022) 50(7):1257–1273 1263

Table 3 Descriptions of the


Metric Name Equation
selected surface metrics
PM1 PN1
Ssk Surface Skewness 1
MNSq3 k¼0 l¼1 ½zðxk ; yl Þ  l3
PM1 PN1
Sku Surface Kurtosis 1
MNSq4 k¼0 l¼1 ½zðxk ; yl Þ  l4
Sa Average roughness R R N1    
M1
1
MN l¼0 z xk; yl  l
K¼0
Svi Valley Fluid Retention Index V v ðh0:80 Þ
ðM1ÞðN1Þdx dy =Sq

where l = the mean height, M = raster rows, N = raster columns, Sq = the Root Mean Square, h = Height

matrix f (x, y). The Fourier transform then becomes subsets in 2013 and 2020 shows that there are small
(Eq. 5): changes in the area and arrangement of urban and agri-
ZZ þ1 cultural patches.
F ðx; vÞ ¼ f ðx; yÞe2piðxxþvyÞ dxdy ð5Þ
1
Accuracy Assessment
where x, v = frequency coordinates.
In general, when Fourier transforms have been applied Table 4 shows the results of the classification accuracy of
to a satellite image, lower frequencies are located in the the two subsets under study. Using ground truth points,
center of the Fourier spectrum, while higher frequencies confusion matrices were generated which indicate the
are located around the center of this spectrum. Hence, it is overall accuracy and kappa coefficient for the classified
expected that images that are homogeneous show high maps. The overall accuracies for the subsets are accept-
values in the center of the x, v scatterplot, and very low able (more than 88%) (Chettry & Surawar, 2021), and the
values around the spectrum. The increase in fragmentation kappa coefficient varies from 0.78 to 0.86, indicating that
leads to higher frequencies, resulting in a more complex we obtained acceptable results.
Fourier spectrum, in which higher values occur in the high-
frequency part of the Fourier spectrum (Rocchini, Metz, Comparison Between Getis Statistics
et al., 2013). Fast Fourier transform function (FFT) based and landscape Metrics
on ENVI5.3 software was calculated to study the landscape
changes. Since Fourier transforms can be performed on one Table 5 shows the results of the landscape metrics calcu-
image only, the images were decomposed into their main lated for the two subsets in the years 2013 and 2020. In this
components using the principal component analysis (PCA). table, seven landscape metrics are presented, each of which
PCA is a linear transformation technique related to factor analyzes a specific aspect of landscape patterns. For
analysis. Given a set of image bands, PCA produces a new example, the PLAND metric shows the percentage of
set of images, known as components that are uncorrelated landscape occupied by a cover. The results of this metric
with one another and are ordered in terms of the amount of show that the cultivated farms increased from 13 to 16% in
variance they explain from the original band set. In this subset 1 and 28% to 34% in subset 2 between the years
process, the first component (PCA1) contains the most 2013 to 2020. According to the PLAND metric, the urban
amount of information and explains 95 to 99 percent of the cover has also increased. PD metric shows that the agri-
total variance in the original data set. Therefore, the first cultural patch density increased in subset 1 but decreased in
component of Landsat images of 2013 and 2020 for both subset 2.
subsets was used as input for the FFT function. The density of urban patches has increased in both
subsets. LPI metric shows the percentage of landscape
occupied by the largest patch of a cover. According to this
Results metric, the area of the largest agricultural patch has
increased in both subsets. Urban patches areas have also
Image Classification increased in both subsets. MESH measures the fragmen-
tation degree of the patches. The high values of this metric
Figure 3 shows the classified maps of the subsets selected indicate the reduction of fragmentation in a landscape.
for this study. This figure shows the three classes, namely Based on this metric, urban and agricultural patches have
agriculture, urban, and other in 2013 and 2020. In this become less fragmented in both subsets. AI metric also
figure, the two land cover classes of agriculture and urban shows the degree of aggregation of patches. In the present
are separated from other covers. A visual comparison of the

123
1264 Journal of the Indian Society of Remote Sensing (July 2022) 50(7):1257–1273

Fig. 3 Land use maps indicating two classes of urban, agriculture. (A1) subset 1 in the year 2013, (A2) subset 1 in the year 2020, (B1) subset 2 in
the year 2013, and (B2) subset 2 in the year 2020

Table 4 Classification accuracy of the land cover map generated by study, this metric does not show any specific changes in the
the maximum likelihood approach for the subsets in different years pattern of patches. High values of the SPLIT metric indi-
Year Subset Overall accuracy kappa coefficient cate an increase in fragmentation in a cover. The results of
this metric also show that the structural pattern of urban
2013 Subset 1 92% 0.80 and agricultural patches in 2020 has become more aggre-
Subset 2 90% 0.84 gated than in 2013.
2020 Subset 1 88% 0.78 As mentioned above, the six landscape metrics were
Subset 2 91% 0.86 calculated using moving windows for comparison with
Getis statistics. Moving window images of PLAND metric
for agriculture class are presented in Fig. 4. High values in

Table 5 Landscape metrics for


Land use PLAND PD LPI MESH AI SPLIT
two subsets understudy in the
years 2013 and 2020 Subset 1–2013 Agri 13 4.7 3.7 4 78.6 552
Urban 10 3.2 4.3 6 79.2 328
Subset 2–2013 Agri 28 5 6.5 21 82.7 110
Urban 15 3.9 5.4 9 80.5 248
Subset 1–2020 Agri 16 5.8 4.9 7 78.3 297
Urban 18 6 12.2 34 77.8 65
Subset 2–2020 Agri 34 4.4 7.2 32 84 74
Urban 16 4.2 5.4 13 82 184

123
Journal of the Indian Society of Remote Sensing (July 2022) 50(7):1257–1273 1265

Fig. 4 Output raster grids of moving window analysis for agriculture class. PLAND metrics derived from land-use maps for (A1) subset 1in the
year 2013, (A2) subset 1 in the year 2020; (B) subset 2in the year 2013; (c) subset 2in the year 2020

these images indicate cultivated farms, and low values are statistically weak correlation with the Getis of NDVI and
related to built-up areas and other covers. The statistical PNBI indexes.
relationship between the images obtained from the moving Results of the density function applied to Getis of NDVI
window and the corresponding images obtained from the between the years 2013 and 2020 for both subsets are
Getis statistics (Figs. 5 and 6) was estimated, and the presented in Fig. 5 (A3 and B3). To obtain the values
results are presented in Table 6. According to this table, the limited to the cultivated farms in the Getis image in subset
statistical relationship between landscape metrics and Getis 1, we masked the corresponding values of the farms from
of NDVI is stronger than Getis of PNBI in both subsets. the Getis image in ArcGIS software. The output image was
Among the landscape metrics, PLAND, LPI, and MESH a raster layer with a range of -5 to 9, so in Fig. 5-A3, this
show the strongest relationship with the Getis index, range was considered as vegetation or cultivated farms.
respectively. Metrics such as PD, AI, and SPLIT showed a Therefore, this plot can be divided into two parts based on

123
1266 Journal of the Indian Society of Remote Sensing (July 2022) 50(7):1257–1273

A1 B1

A2 B2

A3 B3
0.06 0.04
0.035
0.05
Density Proability
Density Proability

0.03
0.04
0.025
0.03 0.02

0.02 0.015
0.01
0.01
0.005
0 0
7- 6- 5- 4- 3- 2- 1- 0 1 2 3 7- 6- 5- 4- 3- 2- 1- 0 1 2 3 5 6 7 8
Ges of NDVI Values Ges of NDVI Values

Fig. 5 Outputs images and change detection plots of Getis of NDVI in change detection plot for subset 1 (B1) Getis of NDVI for subset 2 in
subsets understudy in the years 2013 and 2020. (A1) Getis of NDVI 2013, (B2) Getis of NDVI for subset 2 in 2020, and (B3) change
for subset 1 in 2013, (A2) Getis of NDVI for subset 1 in 2020, (A3) detection plot for subset 2

123
Journal of the Indian Society of Remote Sensing (July 2022) 50(7):1257–1273 1267

A1 B1

A2
B2

A3 B3
0.035 0.03

0.03 0.025
Density Probability

Density Proablity

0.025 0.02
0.02
0.015
0.015
0.01
0.01
0.005
0.005
0
0
3- 2- 1- 0 1 2 3 4
2- 1- 0 1 2 3 4 5 6 14
Ges of PNBI Values Ges of PNBI Values

Fig. 6 Outputs images and change detection plots of Getis of PNBI in change detection plot for subset 1 (B1) Getis of PNBI for subset 2 in
subsets understudy in the years 2013 and 2020. (A1) Getis of PNBI 2013, (B2) Getis of PNBI for subset 2 in 2020, and (B3) change
for subset 1 in 2013, (A2) Getis of PNBI for subset 1 in 2020, (A3) detection plot for subset 2

Getis of NDVI values: (1) the first part between -7 to -5 cross between Getis charts in 2013 and 2020, and from this
that is related to non-agricultural lands including bare soil value onwards, the number of farm cells has increased in
and built-up areas; and (2) the second part between -5 to 9 2020 compared to 2013 which is compatible with the
that is related to cultivated farms. A value of -5 shows a results of landscape metrics. We did the same procedure

123
1268 Journal of the Indian Society of Remote Sensing (July 2022) 50(7):1257–1273

Table 6 Regression statistics of


Regression equation R2 (%) P value
Getis index and landscape
metrics (n = 2500). GN and GP Getis of NDVI* 2013 PLAND = 34.59 ? 6.699 GN 87.6 0.000
are the abbreviations of Getis of
NDVI and Getis of PNBI, LPI = 31.22 ? 6.412 GN 83.6 0.000
respectively PD = 17.15 ? 0.1541 GN 0.3 0.010
MESH = 1.761 ? 0.4701 GN 73.5 0.000
AI = 83.25 ? 1.358 GN 73 0.000
SPLIT = 233.31–31.6 GN 19.4 0.000
Getis of PNBI* 2013 PLAND = -15.74 ? 27.70 GP 26.7 0.000
LPI = -15.01 ? 25.61GP 23.6 0.000
PD = 9.983 ? 4.248 GP 5 0.010
MESH = -1.592 ? 1.942 GP 18 0.000
AI = 84.89–3.135 GP 0.7 0.000
SPLIT = 2087–967.9 GP 12.9 0.000
Getis of NDV 2013 PLAND = 38.84 ? 6.797 GN 95.9 0.000
LPI = 35.21 ? 6.826 GN 92.5 0.010
PD = 17.79–03.343 GN ‘1.7 0.000
MESH = 2.275 ? 0.6215 GN 82.5 0.000
AI = 84.20 ? 1.236 GN 15.1 0.000
SPLIT = 248.8–91.51GN 20.3 0.000
Getis of PNBI 2013 PLAND = -4.747 ? 29.07 GP 35.2 0.010
LPI = -5.711 ? 28.44 GP 33.3 0.000
PD = 13.43 ? 1.069 GP 0.6 0.000
MESH = -1.257 ? 2.624 GP 31 0.000
AI = 78.39 ? 2.874 GP 1 0.000
SPLIT = 1084–40.5 GP 6.5 0.010
Getis of NDVI* 2020 PLAND = 33.07 ? 6.545GN 87.8 0.000
LPI = 29.88 ? 6.378 GN 83.5 0.000
PD = 18.05 ? 0.08573 GN 0.1 0.000
MESH = 1.723 ? 0.484 GN 72.9 0.000
AI = 81.84 ? 1.064 GN 5.4 0.010
SPLIT = 297.0–137.1 GN 21.7 0.000
Getis of PNBI* 2020 PLAND = -14.33 ? 10.40 GP 39.1 0.000
LPI = -15.12 ? 9.846 GP 34.3 0.000
PD = 14.94 ? 0.9994 GP 2.1 0.000
MESH = -1.663 ? 0.7706 GP 26.3 0.010
AI = 73.20 ? 0.8184 GP 0.6 0.000
SPLIT = 1482–254.6 GP 15.8 0.000
Getis of NDVI 2020 PLAND = 42.65 ? 7.114 GN 94 0.000
LPI = 39.04 ? 7.271 GN 91 0.000
PD = 17.93–0.4969 GN 3.6 0.010
MESH = 2.638 ? 0.6995 GN 82.1 0.000
AI = 84.13 ? 1.552 GN 22.6 0.000
SPLIT = 208.1–78.69 GN 19.4 0.000
Getis of PNBI 2020 PLAND = -1.921 ? 10.50 GP 43.7 0.000
LPI = -1.996 ? 10.31 GP 41.4 0.000
PD = 14.47 ? 0.1238 GP 0.1 0.173
MESH = -1.435 ? 0.9160 GP 34 0.000
AI = 77.12 ? 1.384 GP 2.5 0.000
SPLIT = 1308–214.4 GP 13.1 0.000
*
Data belong to subset 1

123
Journal of the Indian Society of Remote Sensing (July 2022) 50(7):1257–1273 1269

for subset 2, which, similar to subset 1, the values of cul- The surface kurtosis (Sku) metric only shows the dom-
tivated fields were between -5 and 9. Therefore, this inance of a cover, and according to this metric, there is a
chart also shows the increase in a farm area in 2020. significant decrease in the cover’s dominance in both
Changes related to urban cover for subset 1 are shown in subsets from 2013 to 2020. However, it is not clear which
Fig. 6-A3. In this figure, the values related to urban cover cover losses the dominance, so this metric should be
are from 0 to 4, and these two numbers have created two interpreted with the surface kurtosis (Sku) metric simulta-
intersection points in this figure which show the growth of neously. Similar to the surface kurtosis (Sku) metric, the
urban cover in 2020. In subset 2, the range of values lim- average roughness (Sa) metric also does not differentiate
ited to the urban cover for the years 2013 and 2020 was among different covers. The results of this metric show that
between -2 and 2. This chart also shows that between 2013 the patch richness in both subsets has increased between
and 2020, the urban cover area has increased which is the years 2013 and 2020. Valley fluid retention index (Svi)
consistent with the results of landscape metrics. is particularly sensitive to occasional deep valleys and not
occasional high peaks. The results of this metric show that
Surface Metrics in both subsets, the negative values (Bare soil and built-up
areas) in the NDVI index have decreased. These negative
Table 7 shows the results of surface metrics values that values or valleys in the PNBI index are cultivated farms,
were calculated for different vegetation indices in different which their dominance has increased in both subsets from
years. Positive values of surface skewness (Ssk) of NDVI 2013 to 2020.
index imply that in subset 1, the bare soil and built-up areas
are more dominant than vegetation covers (cultivated Fourier Transforms
farms). The surface skewness (Ssk) values of PNBI are
negative which means peaks or high values of built-up Figure 7 shows the Fourier spectrums of both subsets in the
areas dominate subset 1. The surface skewness (Ssk) values years 2013 and 2020. As mentioned above, increasing
of the NDVI index decreased from 0.6 to 0.3 in subset 1 fragmentation leads to higher frequencies, which means
from 2013 to 2020, indicating that cultivated farms have that once fragmentation occurred in the landscape, the
increased in the period. The surface skewness (Ssk) values equitability of values increases. According to Fig. 7, the
of the PNBI index decreased from – 4.5 to – 2.5 meaning equitability of values has increased (higher heterogeneity)
that in the year 2020, the dominance of bare soil and built- in the year 2020, because of the fragmentation process or
up areas has reduced. The surface skewness (Ssk) values increased landscape heterogeneity over the subsets mainly
for subset 2 also showed that the area of cultivated farms due to the increasing cultivated farms. The density function
increased significantly from 2013 to 2010, as the domi- plots (A3 and B3) of both subsets confirm the above results
nance of low values has reduced from - 0.21 to 0.04. and show increasing heterogeneity in the year 2020. The
Similar to subset 1, the dominance of negative values (Bare left part of these plots is related to low-frequency values
soil and built-up areas) of PNBI has decreased from 2013 (lower heterogeneity), and the right part is related to high-
to 2020, implying that barren lands have been replaced by frequency values (higher heterogeneity), comparison of
cultivated farms in 2020. these parts shows that in seven years, both subsets have
changed and heterogeneity has increased in the year of
2020 in both subsets.

Table7 Results of calculated surface metrics for the subsets in the


years 2013 and 2020 Discussion
Ssk Sku Sa Svi
The first alternative metric in this study was local Getis that
Subset 1 NDVI of Landsat 2013 0.64 5.3 16.8 0.7
showed a strong statistical relationship with landscape
NDVI of Landsat 2020 0.3 4 21.8 0.6 metrics especially for measuring landscape heterogeneity
PNBI of Landsat 2013 - 4.5 36.7 7.8 1.6 of cultivated farms. Our results on the relationship between
PNBI of Landsat 2020 - 2.5 16 10.7 2.1 local Getis and landscape metrics revealed two general
Subset 2 NDVI of Landsat 2013 - 0.21 2.6 22.6 0.35 results: (1) The relationship between Getis of NDVI and
NDVI of Landsat 2020 0.04 2.2 30 0.23 landscape metrics was stronger than the Getis of PNBI, and
PNBI of Landsat 2013 - 2.8 21 11.3 0.98 (2) Getis of NDVI had a strong statistical relationship with
PNBI of Landsat 2020 - 1.5 8.5 15.7 1.4 some landscape metrics (PLAND, LPI, and MESH), and at
the same time, it has a weak relationship with others (AI,
PD, SPLIT). A comparison between local Getis and

123
1270 Journal of the Indian Society of Remote Sensing (July 2022) 50(7):1257–1273

A1 B1

A2 B2

A3 B3
0.016 0.016
0.014 0.014
0.012 0.012
Density Probablity
Density Probablity

0.01 0.01
0.008 0.008
0.006 0.006
0.004 0.004
0.002 0.002
0 0
0 40 80 120 160 200 240 0 40 80 120 160 200 240
Fourier of PCA1 Values Fourier of PCA1 Values

Fig.7 Results of Fourier transform applied to PCA1 of subsets under blue color (out of spectrum) indicates low frequency, green color
study. (A1) Fourier spectrum of subset 1 in 2013, (A2) Fourier indicates medium frequency, and yellow color indicates the high
spectrum of subset 1 in 2020, (A3) change detection plot for subset frequency at the center of the spectrum. Visual comparison of these
1,(B1) Fourier spectrum of subset 2 in 2013, (B2) Fourier spectrum of images shows lower values (yellow) in the high-frequency part of the
subset 1 in 2020, and (B3) change detection plot for subset 2. The Fourier spectrum of 2017

landscape metrics has been reported by Fan and Myint PNBI, implying that local spatial autocorrelation indices
(2014). Our results in this part are similar to Fan and Myint are sensitive to input data.
(2014) as they also showed that Getis of NDVI had a One reason for the stronger relationship of Getis of
stronger relationship with landscape metrics than Getis of NDVI and landscape metrics is that the NDVI index

123
Journal of the Indian Society of Remote Sensing (July 2022) 50(7):1257–1273 1271

merely reflects vegetation cover, but in the PNBI index, in patterns (Kedron et al., 2018). In contrast to landscape
addition to urban areas, barren lands also receive high metrics, surface metrics cannot determine the number of
values. Another reason is related to the accuracy of clas- changes in units of area. The advantages of these metrics
sification of urban areas because these areas are composed are that (1) they are sensitive to small land-use changes, (2)
of different human-made covers and have different reflec- they are suitable for examining long-term land-use chan-
tions and structural features; then, processes of urban ges, and (3) they distinguish different covers using domi-
classification become more difficult and more error-prone nance metrics. Frazier and Kedron (2017) point to the
(Rocchini, Foody, et al., 2013). For example, in the process limitations of landscape metrics introduce surface metrics
of image classification, scattered trees or small green space as alternatives and acknowledge that alternative approa-
patches are usually not classifiable with Landsat images or ches are suffering from the issue of correlation and
classified as single pixels that the user removes or merges redundancy. Similar to Getis statistics, Fourier transforms
in urban cover. However, autocorrelation methods use the also showed landscape pattern changes via change detec-
NDVI index, which these scattered trees and small green tion plots, which is an advantage of the method, but it does
spaces are considered in the output images. not provide any information about the type and the amount
Fan and Myint (2014) also found that PLAND and LPI of changed cover. Comparison of Fourier transforms and
metrics showed the highest statistical relationship with Getis plots showed that local Getis was more powerful than
Getis of NDVI (R2 = 95% and 82%, respectively) that is Fourier transforms to detect landscape changes.
confirmed by our results as we also found the strongest
relationship between Getis of NDVI and PLAND and LPI
metrics (R2 = 95% and 92%, respectively). The Getis Conclusions
statistics has also been used for measuring landscape pat-
terns in a change detection scenario (Nolè et al., 2014) and Overall, the use of discrete and continuous approaches
can well differentiate the urban and vegetation. Consistent depends on factors like the scale of the study, time and
with landscape metrics, change detection plots of Getis of budget available, desired ecological processes, and the
NDVI and PNBI showed that cultivated farms and urban degree of heterogeneity of the landscape of interest (Gus-
areas increased between 2013 and 2020 in both subsets. tafson, 2019). However, our results showed that similar to
Generally, the advantages of spatial autocorrelation indices landscape metrics, Getis statistics, surface metrics, and
are; (1) these statistics can distinguish small changes in Fourier transforms also have their disadvantages for mea-
land use and (2) they provide a map as their outputs (3) suring landscape changes. For example, the mathematical
they are suitable for quick evaluation of landscape patterns context and concept of continuous metrics are more com-
because they use images such as NDVI as their input. The plicated than landscape metrics, especially for surface
disadvantage of these criteria is that (1) they do not cal- metrics. In contrast to landscape metrics, alternative
culate land-use changes in area units and (2) they are more methods have no functional metrics for measuring func-
suitable for studying vegetation and urban environments. tional or ecological processes. Using surface metrics, we
Surface metrics were another alternative metric for can estimate changes of land covers between different
quantifying landscape changes that use continuous data as years, but the location of changes cannot be well deter-
the input of their calculation. Results of these metrics are mined because these metrics do not provide a map. Fourier
consistent with Getis statistics and landscape metrics as transforms can show the amount of change but do not
these metrics showed that cultivated farms and built-up provide any information about which cover in what part of
areas increased from 2013 to 2020 in both subsets. Surface the landscape has changed. However, the results of con-
metrics have not been used to measure changes over time, tinuous metrics are quantitative which is very important in
but this important was done in the current study and results the field of landscape ecology. Continuous metrics also
showed the efficiency of these metrics in temporal mea- have high sensitivity to small changes, implying that they
suring landscape changes. These metrics have been used in are more suitable than landscape metrics for detecting
a few studies so far (McGarigal et al., 2009; Mondal, small changes.
2011). Although the results of earlier studies confirm the Calculating continuous metrics is less time-consuming
ability of surface metrics in quantifying landscape hetero- than landscape metrics because they use vegetation indexes
geneity (Frazier & Kedron, 2017), these metrics also have as input that are easy to calculate. The output images from
their weaknesses. For instance, surface metrics are sensi- Fourier analysis and spatial autocorrelation indices provide
tive to the input data like NDVI and PNBI indices, as they visual interpretation especially for Getis statistics that help
estimated different levels of landscape dominance. The us to find locations of clusters. Therefore, continuous
interpretation of surface metrics is also challenging and not metrics have unique characteristics for measuring land-
all surface metrics are suitable for analyzing landscape scape patterns and drawing their changes. We recommend

123
1272 Journal of the Indian Society of Remote Sensing (July 2022) 50(7):1257–1273

the use of continuous metrics for quick assessment of International Journal of Applied Earth Observation and Geoin-
landscape, especially when seeking to detect changes over formation, 31, 154–166.
Fan, C., & Myint, S. (2014). A comparison of spatial autocorrelation
time in large landscapes. We also suggest applying con- indices and landscape metrics in measuring urban landscape
tinuous metrics instead of using landscape metrics when fragmentation. Landscape and Urban Planning, 121, 117–128.
the spatial arrangement of landscape elements is very Forman, R. T. (2014). Land mosaics: The ecology of landscapes and
heterogeneous. Landscape metrics were the simplest met- regions (1995). Berlin: Springer.
Fourier, J. (1822). Theorie analytique de la chaleur, par M. Fourier,
rics used in the present study, which were much simpler Chez Firmin Didot, père et fils.
than other metrics in terms of computational complexity Frazier, A. E. (2016). Surface metrics: Scaling relationships and
and formula. The advantages of all the metrics presented downscaling behavior. Landscape Ecology, 31(2), 351–363.
are: (1) they provided a quantitative description of the Frazier, A. E., & Kedron, P. (2017). Landscape metrics: Past progress
and future directions. Current Landscape Ecology Reports, 2(3),
landscape patterns; and (2) their calculation was easy using 63–72.
existing software in the field of landscape ecology. How- Getis, A., & Ord, J. K. (1992). The analysis of spatial association by
ever, the output of these metrics is expressed as a number use of distance statistics. Geographical Analysis, 24(3),
and not an image, which is a significant weakness of these 189–206.
Gustafson, E. J. (2019). How has the state-of-the-art for quantification
metrics. We emphasize that landscape metrics have some of landscape pattern advanced in the twenty-first century?
drawbacks for measuring landscape patterns and some of Landscape Ecology, 34(9), 2065–2072.
them can be resolved by continuous metrics. Since alter- Hagen-Zanker, A. (2016). A computational framework for general-
native metrics also have disadvantages, discrete and con- ized moving windows and its application to landscape pattern
analysis. International Journal of Applied Earth Observation
tinuous metrics can complement each other in measuring and Geoinformation, 44, 205–216.
landscape patterns. Kedron, P., Zhao, Y., & Frazier, A. E. (2019). Three dimensional
(3D) spatial metrics for objects. Landscape Ecology, 34(9),
2123–2132.
Funding There are no financial conflicts of interest to disclose. Kedron, P. J., Frazier, A. E., Ovando-Montejo, G. A., & Wang, J.
(2018). Surface metrics for landscape ecology: A comparison of
landscape models across ecoregions and scales. Landscape
Availability of Data and Material Data are available on request from
Ecology, 33(9), 1489–1504.
the authors only based on logical requests.
Kulhanek, K., Steinhauer, N., Rennich, K., Caron, D. M., Sagili, R.
R., Pettis, J. S., Ellis, J. D., Wilson, M. E., Wilkes, J. T., &
Code Availability Not applicable. Tarpy, D. R. (2017). A national survey of managed honey bee
2015–2016 annual colony losses in the USA. Journal of
Apicultural Research, 56(4), 328–340.
Declaration Lanorte, A., Danese, M., Lasaponara, R., & Murgante, B. (2013).
Multiscale mapping of burn area and severity using multisensor
Conflict of interest On behalf of all authors, the corresponding author satellite data and spatial autocorrelation analysis. International
states that there is no conflict of interest. Journal of Applied Earth Observation and Geoinformation, 20,
42–51.
Ethics Approval Not applicable. Lausch, A., Blaschke, T., Haase, D., Herzog, F., Syrbe, R.-U.,
Tischendorf, L., & Walz, U. (2015). Understanding and quan-
Consent to Participate All authors consent to participate in writing tifying landscape structure–A review on relevant process char-
this paper. acteristics, data models and landscape metrics. Ecological
Modelling, 295, 31–41.
Consent for Publication All authors consent to publish this paper. Li, H., & Wu, J. (2004). Use and misuse of landscape indices.
Landscape Ecology, 19(4), 389–399.
Liu, D., Hao, S., Liu, X., Li, B., He, S., & Warrington, D. (2013).
Effects of land use classification on landscape metrics based on
References remote sensing and GIS. Environmental Earth Sciences, 68(8),
2229–2237.
Anselin, L. (1995). Local indicators of spatial association—LISA. McGarigal, K., & Cushman, S. (2005). The gradient concept of
Geographical Analysis, 27(2), 93–115. landscape structure. In J. Weins & M. Moss (Eds.), Issues and
Bolliger, J., Wagner, H. H., & Turner, M. G. (2007). Identifying and perspectives in landscape ecology (pp. 112–119). Cambridge:
quantifying landscape patterns in space and time. In A Changing Cambridge University Press.
World, Springer, pp. 177–194. McGarigal, K., Cushman, S. A., Neel, M. C., & Ene, E. (2002).
Buyantuyev, A., Wu, J., & Gries, C. (2010). Multiscale analysis of the FRAGSTATS: spatial pattern analysis program for categorical
urbanization pattern of the Phoenix metropolitan landscape of maps.
USA: Time, space and thematic resolution. Landscape and McGarigal, K., Tagil, S., & Cushman, S. A. (2009). Surface metrics:
Urban Planning, 94(3–4), 206–217. An alternative to patch metrics for the quantification of
Chettry, V., & Surawar, M. (2021). Urban sprawl assessment in eight landscape structure. Landscape Ecology, 24(3), 433–450.
mid-sized indian cities using RS and GIS, Journal of the Indian Metrology, I. (2002). The scanning probe image processor, SPIPTM,
Society of Remote Sensing 1–20. user’s and reference guide version 4.
Cockx, K., Van de Voorde, T., & Canters, F. (2014). Quantifying
uncertainty in remote sensing-based urban land-use mapping.

123
Journal of the Indian Society of Remote Sensing (July 2022) 50(7):1257–1273 1273

Mondal, P. (2011). Quantifying surface gradients with a 2-band Rocchini, D., Foody, G. M., Nagendra, H., Ricotta, C., Anand, M.,
Enhanced Vegetation Index (EVI2). Ecological Indicators, He, K. S., Amici, V., Kleinschmit, B., Förster, M., &
11(3), 918–924. Schmidtlein, S. (2013). Uncertainty in ecosystem mapping by
Nolè, G., Lasaponara, R., Lanorte, A., & Murgante, B. (2014). remote sensing. Computers & Geosciences, 50, 128–135.
Quantifying urban sprawl with spatial autocorrelation techniques Rocchini, D., Metz, M., Ricotta, C., Landa, M., Frigeri, A., & Neteler,
using multi-temporal satellite data. International Journal of M. (2013). Fourier transforms for detecting multitemporal
Agricultural and Environmental Information Systems (IJAEIS), landscape fragmentation by remote sensing. International Jour-
5(2), 19–37. nal of Remote Sensing, 34(24), 8907–8916.
Park, Y., & Guldmann, J.-M. (2020). Measuring continuous landscape Shao, G., & Wu, J. (2008). On the accuracy of landscape pattern
patterns with Gray-Level Co-Occurrence Matrix (GLCM) analysis using remote sensing data. Landscape Ecology, 23(5),
indices: An alternative to patch metrics? Ecological Indicators, 505–511.
109, 105802.
Rahimi, E., Barghjelveh, S., & Dong, P. (2021). Quantifying how Publisher’s Note Springer Nature remains neutral with regard to
urban landscape heterogeneity affects land surface temperature jurisdictional claims in published maps and institutional affiliations.
at multiple scales. Journal of Ecology and Environment, 45(1),
1–13.

123

You might also like