IOP Conference Series: Materials Science and Engineering

PAPER • OPEN ACCESS

Accuracy assessment of supervised classification methods for extraction

land use maps using remote sensing and GIS techniques
To cite this article: Zahraa Abbas and Hussein Sabah Jaber 2020 IOP Conf. Ser.: Mater. Sci. Eng. 745 012166

View the article online for updates and enhancements.

This content was downloaded from IP address 158.46.184.245 on 21/03/2020 at 17:24

The Fourth Postgraduate Engineering Conference IOP Publishing
IOP Conf. Series: Materials Science and Engineering 745 (2020) 012166 doi:10.1088/1757-899X/745/1/012166

Accuracy assessment of supervised classification methods for

extraction land use maps using remote sensing and GIS
techniques

Zahraa Abbas1, Dr. Hussein Sabah Jaber2

1-2
College of Engineering, University of Baghdad, Iraq
E-mail: [email protected]

Abstract. Remotely sensed imagery identifying as the best type of data has information
throughout the world. The imagery has a importance information, since it can show up-date-to
day information, and provide a truly information. Different kinds of classifiers have been used
to perform that. However, there is no once test for Land cover and Land use in Hilla city. The
study aims to create land use classification by making a comparison between different
algorithms in Hilla city, Babylon, Iraq. The WorldView-2 imagery is used to perform the per-
processing, analysing of our comparison. Under the steps of pre-processing, the several
corrections were made and performed on the imagery. For processing stages, two approaches
were used; (1) Support Vector Machine and (2) Maximum Likelihood. The result reveals, that
the Support Vector Machine method has the most significant of overall accuracy equal to
94.48% with kappa coefficient equal to 0.90, and these values much better and higher than
those of Maximum Likelihood algorithm in estimating and extracting of Land cover/Land use.
Therefore, this algorithm has been suggested to be applied as an optimal classifier for
extraction of land use maps due to its higher accuracy and better consistency within the study
area.

1. Introduction
Remote sensing and Geographic information system have capability of detecting and/ or monitoring
the features of earth’s surface using satellite images have different radiometric, spectral, spatial, and
temporal resolution. These technologies resolutions have several advantages in order to minimize time
and cost to extract Land cover and Land use (LULC) [1, 2]. In addition, the remote sensing provides
important and different kinds of remotely data source to extract LULC information. Remote sensing
data are widely used and applied to perform classification of LULC throughout the world [2]. Remote
sensing data has the ability of updating information about all the features that locate on earth’s surface
[3]. There are different satellite systems provides different imagery has different spatial resolutions
applied for LULC detection such as WORLDVIEW-2, Quick Bird, Landsat, Spot and so on [1-3]. The
classification of remote sensing imagery is an important method in order to determine the LULC
information [2]. Classification approaches divided into two categories: pixel base and /or objected
oriented base classifiers [3]. The Maximum Likelihood approach (MLC) widely applied for LULC
extraction, by creating decisions of surface regarding of the covariance and mean of each single class
[4]. However, the non-parametric technique Support Vector Machine (SVM) classifier, perform the
classification by no assumptions and separate the features on each class [5, 6]. SVM is a non-
parametric algorithm, which contains a series of learning algorithms are conducted for classes
classification and regression [5-7]. The SVM classification was proposed by Vapnik and

Content from this work may be used under the terms of the Creative Commons Attribution 3.0 licence. Any further distribution
of this work must maintain attribution to the author(s) and the title of the work, journal citation and DOI.
Published under licence by IOP Publishing Ltd 1
The Fourth Postgraduate Engineering Conference IOP Publishing
IOP Conf. Series: Materials Science and Engineering 745 (2020) 012166 doi:10.1088/1757-899X/745/1/012166

Chervonenkis [8] and after that discussed in more detail by Vapnik [9]. Add to that, the SVM ws
widely applied in different remote sensing projects and applications [10, 11]. Huang et al. [12] applied
the SVM classifier for a LULC classification by using Landsat thematic mapper (TM) imagery. The
SVM classification has a good accuracy and superior to the accuracy of using different methods such
as (1) maximum likelihood and (2) a decision tree [1-3]. Yousefi et al. [13] has tested a nine different
supervised classification approaches and they are (neural network, Spectral angle mapper,
maximum likelihood, SVM, mahalanobis distance, binary code, minimum distance, spectral
information divergence and parallelepiped) for LULC mapping in province of Mazandaran, Iran
and they used the imagery of Landsat ETM+. The results showed that the SVM classifier has the best
accuracy than the other classifiers. The main objective of this study is to perform LULC classification
using MLC and SVM approaches in Hilla city, Babylon, Iraq, and make comparison to find out the
perfect algorithm to classify the study area.
2. Materials and methods
The detection technique to extract the thematic map of LULC was tested and referenced by previous
work [12]. In this research, WorldView-2 satellite system imagery of 2017 was used to perform the
pre-processing and processing steps. In this research, geometric, radiometric, and atmospheric
corrections were applied to remove image noise. Then, layer stacking and image sub-setting were
conducted.
The next step was to applied different kinds of supervised classification algorithms (MLC and
SVM) that were applied to the collected imagery to perform imagery classification to map the LULC
theme of Al-Hilla city, Babylon, Iraq. The accuracy assessment and statistical analysis of the expected
results were conducted to get an accurate and suitable approach to classify the study area of this
research. Figure 1 reveals the flowchart of our methodology.

Figure 1. The adopted Flowchart of the study area methodology.

2.1. Study Area Location and Descriptions

Hilla city has coordinates of 32°29′N 44°26′E, with elevation 112 ft equal to 34 m. It is around 100 km
(62 miles) to the south of Baghdad. The Hilla city relates to the governorate of Babylon and it is the

2
The Fourth Postgraduate Engineering Conference IOP Publishing
IOP Conf. Series: Materials Science and Engineering 745 (2020) 012166 doi:10.1088/1757-899X/745/1/012166

capital of Babil province, and it is located in central Iraq. Its boundaries area surrounding with other
governorates, they are Baghdad, Kerbala, Anbar, Najaf, Wassit, and Qadissiya. It has Euphrates River,
it is one of biggest rivers in Iraq, and it intersects its governorate and divided into the AL- Hindiyah
and Al- Hilla branches south of the Musayib town, to create a network of canals runs through the lands
of governorate to use as an irrigation system and supply the regions, fields and farms with water. Hilla
city has a population of around 970, 000 people regarding registration of 2018. Figure 2 shows
Location of the Hilla city.

Figure 2. The study area location map.

2.2. The Used Date

A World View-2 Satellite system is one of a large number of systems that provide an imagery which
has high-resolution multispectral and panchromatic bands. The WorldView-2 Satellite has one
panchromatic band, with eight multispectral bands, four of these bands have standard colors (red (R),
green (G), blue (B) and near-infrared1 (NIR1) with another four new bands and they are; yellow, red
edge, coastal, and near-infrared 2. The WorldView-2 Satellite images have a full-color for different
kinds of applications such as exploration, enhanced spectral analysis, disaster relief, mapping,
monitoring applications, defense, intelligence applications, land-use/land cover planning, simulation
environments and visualization . The system WorldView-2 can act like a paintbrush, sweeping from
back to forth in order to capture large areas of multispectral images in one single pass. WorldView-2
satellite system captures around 1 million km2 every single day. Table 1and Figures 3 and 4 show the
WorldView-2 satellite system and its bands.

3
The Fourth Postgraduate Engineering Conference IOP Publishing
IOP Conf. Series: Materials Science and Engineering 745 (2020) 012166 doi:10.1088/1757-899X/745/1/012166

Figure 3. The WorldView-2 satellite system bands.

Figure 4. The WorldView-2 image of the study area.

Table1. The specifications of WorldView-2 satellite system.
date of launch 8/10/2009
altitude of orbit 770 kilometers
orbit type sun synchronous
orbit period 100 mints.
sensor bands Panchromatic band and 8 multi-spectral bands
ground sample distance panchromatic= 0.46 m GSD with nadir
and 0.52 m GSD with 20° when sensor is off-nadir
sensor resolution GSD
multispectral bands = 1.84 m GSD with Nadir, 2.4 m GSD with
20° off-nadir
swath width 16.4 Km at nadir
attitude determination and control 3-Axis Stabilized
GPS position accuracy < 500 m
1.1 days at 1 meter GSD or less 3.7 days at 20° off-nadir or less
revisit frequency
(0.52 meter GSD)
geolocation accuracy demonstrated <3.5 m ce90 without GCPs
(Source: https://www.satimagingcorp.com/satellite-sensors/worldview-2/)

4
The Fourth Postgraduate Engineering Conference IOP Publishing
IOP Conf. Series: Materials Science and Engineering 745 (2020) 012166 doi:10.1088/1757-899X/745/1/012166

3. Imagery Processing and Analysis

3.1. Atmospheric, Radiometric and Geometric Corrections
Atmospheric, radiometric and geometric corrections are required for the remove the imagery noise to
perform different applications of remote sensing such as; image classification [11]. The reason is that
atmospheric correction of a date image would often mean to remove a values from imagery pixels
from the spectral bands of the satellite images [13]. However, the geometric correction of WorldView-
2 to a satellite image is conducted by collecting ground control points (GCPs) from the fieldwork
method using Envi 5.2 software to perform the geometric correction.

3.2. Fieldwork
The fieldwork conducted by using Handheld GPS type Garmin 78s, it is one of the most popular
devices have been used to collect the locations of points locate in the study area with different kind of
remote sensing applications. The accuracy of this device is for horizontal < 10m, and for vertical =
0.05 m/s steady-state. Figure 5 shows the used Garmin GPS.

Figure 5. Handheld GPS type Garmin 78s,

10 GCPs were collected from fieldwork to perform the geometric correction and to choose a
suitable projection to the used image which is the UTM, in Zone 38N with WGS 84 datum and table 2
includes the collected GCPs from fieldwork, the geometric correction performed using Envi software
as indicates in Figure 6.

Table 2. The 10 Collected GCPs from fieldwork of the study area.

No. Longitude (E) Latitude (N) Features
1 44°25'15.03" 32°29'55.17" Bridge
2 44°26'12.82" 32°29'24.17" Round about
3 44°27'26.65" 32°29'47.85" Al-Bakerly inter section
4 44°25'09.14" 32°26'47.22" Nader inter section
5 44°25'31.78" 32°28'03.38" Mother square
6 44°26'22.21" 32°29'02.56" Province bridge
7 44°23'41.21" 32°26'40.16" Road inter section
8 44°24'48.45" 32°26'02.35" Factory gate
9 44°26'21.60" 32°26'14.92" Small bridge
10 44°24'01.47" 32°28'37.69" Al-Asatetha road inter section

5
The Fourth Postgraduate Engineering Conference IOP Publishing
IOP Conf. Series: Materials Science and Engineering 745 (2020) 012166 doi:10.1088/1757-899X/745/1/012166

In addition, the authors use the Nearest Neighbour approach is used for resampling uncorrected
pixels. Finally, a root mean square error (RMSE) of images is used to obtain to check out their
accuracy and we got less than 0.4 pixels, and this value is acceptable [14]. Figure 7 shows the imagery
after perform all corrections and it became a noise-free image.

Figure 6. Geometric correction of WorldView-2 satellite image by using GCPs.

Figure 7. The imagery after performing all correction and calibration.

3.3. Image Classification
Image classification was carried out by using the MLC and SVM algorithms. In the following
subsections, brief explanations of the two algorithms are provided.
3.3.1. Selecting Training and Testing Sites
After removing all imagery noise by radiometric and geometric correction and conduct image clipping
and layer stacking, the WorldView-2 satellite image becomes ready for further digital image

6
The Fourth Postgraduate Engineering Conference IOP Publishing
IOP Conf. Series: Materials Science and Engineering 745 (2020) 012166 doi:10.1088/1757-899X/745/1/012166

processing. The study area was classified into four major classes to perform different classification
techniques based on the most located features into the study area to obtain the thematic maps of land
cover and land cover. Four classes are (Water bodies, Soil, Urbanization area and Vegetation area)
collected training sites were conducted by using Envi software be selected polygons of the region of
interest (ROIs) for each class, then used these ROIs into classification to generate the thematic map of
Land use and land cover for the study area. For this stage was about collecting the testing sites to study
area, these sites very important to determine the accuracy assessment for each classification algorithms
and to check the validation, producer and user accuracy of classifications. Usually testing sites are
ground truth samples, and the better collecting from the field of the study area to get pure ground truth
samples. Collecting the testing sites for this study was conducted by using office work from the
imagery. Figure 8 (a &b) shows the number of pixels selected to be as training and testing sites for
each class of the study area.

Figure 8. (a &b). Show training and testing sites for each class of our study area
3.3.2. Maximum Likelihood Approach
A maximum-likelihood algorithm is one of the wide classifiers has been used for supervised image
classification in different remote sensing application [14]. Erdas [15] mentioned that this approach
works by computing weighted distances and/or likelihood D relate to an unknown vector X belong to
known classes. However, Mc is regarding Bayesian equation [15]:

D=ln(ac)- [0.5ln(ǀCovcǀ)]-[0.5(X-Mc)T(Covc-1)(X-Mc) (1)

Where, C is a particular class, ac refers to the percent probability for any candidate pixel of
class C. when Covc refers to pixels covariance matrix in class c. However, ǀCovcǀ is a Covc
determinant, (Covc-1) inverse of Covc, the T refers to transposition function [15]. After that the
results is revealed and the validation is done by using the confusion matrix. The results of applied
MLC and the producer and user accuracies are presented in Table 3 below for each class. Figure 9
WorldView-2 satellite image after perform MLC classification.
Table 3. Producer and user accuracies of MLC classification
Class Prod. Acc. (%) User Acc. (%)
Water Bodies 99.49 79.81
Vegetation 82.70 99.57
Urbanization 89.32 50.57
Soil Area 74.89 95.10

7
The Fourth Postgraduate Engineering Conference IOP Publishing
IOP Conf. Series: Materials Science and Engineering 745 (2020) 012166 doi:10.1088/1757-899X/745/1/012166

Figure 9. LULC thematic map with MLC classification.

3.3.3. Support Vector Machine Algorithm
SVM has been widely applied for imagery classification in the remote sensing applications
[12]. The SVM aims to reach the optimal separating of hyper plane and/or hyper plane that
locates in high and/or dimensional space, in order to locate an optimal boundary between classes.
The approach this method was proposed by the authors Vapnik and Chervonenkis [8] and later
discussed by the author Vapnik [9]. The success of this method regarding to how it is good to
process its trains [16, 17]. With Envi. V-5.2 done the SVM classification, this classifier based on also
pixels of imagery that mean conducted segmentation first for whole image [17]. The kernel type is
radial based function, there were more than 50 sample selected for each class in training sites. After
that the results is revealed and the validation is done by using the confusion matrix. The results of
applied SVM and the producer and user accuracies are presented in Table 4 for each class. Figure 10
shows kernel as a radial based function, Figure 11 reveals WorldView-2 satellite image after perform
SVM classification.

8
The Fourth Postgraduate Engineering Conference IOP Publishing
IOP Conf. Series: Materials Science and Engineering 745 (2020) 012166 doi:10.1088/1757-899X/745/1/012166

Figure 10. Kernel type as a radial based function of SVM classification.

Table 4. Producer and user accuracies.

Class Prod. Acc. (%) User Acc. (%)

Water Bodies 96.23 92.25
Vegetation 94.28 97.34
Urbanization 94.81 69.18
Soil Area 87.99 96.95

Figure 11. LULC thematic map with SVM classification.

9
The Fourth Postgraduate Engineering Conference IOP Publishing
IOP Conf. Series: Materials Science and Engineering 745 (2020) 012166 doi:10.1088/1757-899X/745/1/012166

4. Classification Accuracy Assessment

The final step of the satellite image classification involves the accuracy assessment stage [18].
Accuracy assessment is a quantification of estimation with aid of remotely sensed dataset to
classification conditions and it is useful for evaluation of classification approach, and it is also
important to determination the error that might be involves. The accuracy assessment of our
classifiers is proposed in the form confusion matrix [19].In this research, to validate the our
classification results with the testing sites that mention in section (3.3.1) above, with the help of
collected GCPs from fieldwork (that collected by using Handheld Garmin 78s GPS) were chosen
randomly for performing the accuracy assessment of each classification that made by SVM and
MLC.
5. Results and discussion
The overall accuracy of the SVM and ML were 94.48% and 88.30%, respectively. However, kappa
coefficients of both SVM and ML were 0.90and 0.80, respectively. Table 5 and Figures 12 & 13 show
the comparison between the Overall Accuracy and Kappa coefficient for each classification of LULC.
Table 5. Comparison between the classifications accuracy for land cover.
No. Approaches Overall Accuracy (%) Kappa coefficient
1 Maximum likelihood 88.30 0.80
2 Support vector machine 94.48 0.90

Figure 12. The overall accuracy of two classifiers

According to the above figures, it clearly seems that the SVM classifier has the highest results of
overall accuracy and the kappa coefficients than those of the MLC approach for our study area in Hilla
city, Babylon, Iraq. Moreover, both Huang et al. [12] and Otukei & Blaschke [22] are evaluated
various methods for classifications in LULC mapping, and they were revealed that the results of
SVM approach compared to the results of ML algorithm and also to decision trees method has the
most highest accuracy assessment for mapping the LULC. The author Deilmai et al. [23-27] in his
research, compared between 2 classification algorithms SVM and MLC in order to extract LULC
thematic map in Johor city, Malaysia., his results revealed the results of SVM classifier regarding to
kappa coefficient was equal to 0.86 and it was most an accurate algorithm.

10
The Fourth Postgraduate Engineering Conference IOP Publishing
IOP Conf. Series: Materials Science and Engineering 745 (2020) 012166 doi:10.1088/1757-899X/745/1/012166

Figure 13. The kappa coefficient of two classifiers.

6. Conclusions
Nowadays, applying LULC mapping is a critical issue for collecting information for master city
planning, detecting and monitoring the environment. Using remote sensing technology with different
satellite imagery is recommended as the most perfect strategy for LULC mapping. Using images of
different systems, the limitation and capability of these images can be detected. The purpose behind
this research is to generate the LULC thematic maps by performing a comparison between two
algorithms, by using both of SVM and MLC methods in Hilla city, Babylon, Iraq, by using imagery of
WorldView-2 satellite system. After performing all necessary corrections by pre-processing,
processing and analysing stages on the satellite images, two different approaches were used to
classify our satellite image. The validation results reveal, that the results of SVM algorithm with
overall accuracy of 94.48 % and with a kappa coefficient of 0.90, it is a higher accuracy if we compare
to that one that got from applied the MLC method for LULC mapping. The SVM approach is
suggested to be as the best classifier to extract the maps LULC, simply because, it has higher
accuracies for our study area. In this research, we confirm that the proper performance of SVM
method with the extracted LULC thematic map can can be facilitated a management in level with
sustainable developing.

References
[1] H Dibs and S AL-Hedny. 2019. Detection Wetland Dehydration Extent With Multi-Temporal
Remotely Sensed Data Using Remote Sensing Data Using Remote Sensing Analysis and GIS.
IJCIET, 2019. 10 (1)143–154.
[2] H Dibs. 2018a. Comparison of Derived Indices and Unsupervised Classification For Al-Razaza
Lake Dehydration extent Using Multi-Temporal Satellite Data and Remote Sensing Analysis ARPN J
Eng Appl Sci 13 (24).
[3] H Dibs, S AL-Hedny, H S Karkoosh. 2018b. Extraction Detailed Buildings 3D Model With
Using High Resolution Satellite Imagery By Remote Sensing and GIS Techniques AL-Qasim Green
University: A Case Study IJCIET 9 (7) 1097–1108.
[4] J Richards & X Jia. Remote sensing digital image analysis. 4th edition. Heidelberg, springer,
Germany. 2006. 359-88.
[5] M Bray & D Han. Identification of Support Vector Machines for Runoff Modelling. J.
Hydroinform. 2004, 6, 265-80.

11
The Fourth Postgraduate Engineering Conference IOP Publishing
IOP Conf. Series: Materials Science and Engineering 745 (2020) 012166 doi:10.1088/1757-899X/745/1/012166

[6] A Moghaddamnia M Gosheh M Nuraie M Mansuri D Han & E Schmitter. Performance

Evaluation of LLR, SVM, CGNN and BFGSNN Models to Evaporation Estimation. Water Geosci.
2010, 9, 108-13.
[7] D Han L Chan and N Zhu. Flood Forecasting Using Support Vector Machines. J.
Hydroinform. 2007, 9, 267-76.
[8] N Vapnik & A Chervonenkis. On Tthe Uniform Convergence of The Relative Frequencies of
Events to Their Probabilities. Theory Probability. Appl. 1971, 16, 264-80.
[9] N Vapnik. An Overview of Statistical Learning Theory. IEEE Trans. Neural Network. 1999, 10,
988-99.
[10] B Dixon & N Candade. Multispectral Land Use Classification Using Neural Networks
and Support Vector Machines. International Journal of Remote Sensing, 2008, 29. 1185-206.
[11] G Foody & A Mathur. Toward Intelligent Training of Supervised Image Classifications:
Directing Training Data Acquisition for SVM Classification. Remote Sensing and Environment. 2004,
93, 107-17.
[12] C Huang LS Davis and JRG Townshend. An Assessment of Support Vector Machines for Land
Cover Classification. Int. J. Remote Sens. 2002, 23, 725-49.
[13] S Yousefi M Tazeh S Mirzaee HR Moradi and SH Tavangar. Comparison of
Different Classification Algorithms in Satellite Imagery to Produce Land Use Maps, Case
study: Noor city. J. Appl. rs gis tech. Nat. Res. Sci. 2011, 2, 15-25.
[14] C Tucker D Grant & J Dykstra. NASA’s Global Orthorectified Landsat Data set. Photogramm.
Eng. Rem. Sens. 2004, 70, 313-22.
[15] Erdas Inc. Erdas field guide. Erdas Inc., Atlanta, Georgia, 1999.
[16] EE Osuna & R Freud. Support Vector Machines: training and applications.
Massachusetts institute of technology cambridge, USA, 1977. p. 1-144.
[17] A Erener. Classification Method, Spectral Diversity, Band Combination and Accuracy
Assessment Evaluation for Urban Feature Detection. Int. J. Appl. Earth Observ. Geoinform. 2013, 21,
397-408.
[18] B Manisha G Chitra & N Umrikar. Image Classification Tool For Land Use/ Land Cover
Analysis: A Comparative Study of Maximum Likelihood and Minimum Distance Method. Int. J.
Geo. earth. environ. 2012, 6, 189-96.
[19] R Congalton. A Review of Assessing The Accuracy of Classifications of Remotely
Sensed Data. Remote Sens. Environ.1991, 37, 35-46.
[20] D Lu P Mausel E Brondizio & E Moran. Change Detection Techniques. Int. J. Remote Sens.
2004, 25, 2365-407.
[21] K Sundarakumar M Harika S Aspita S Yamini & K Balakrishna. Land Use and Land
Cover Change Detection and Urban Sprawl Analysis of Vijayawada City Using Multi-temporal
Landsat data. Int. J. Eng. Sci. Tech. 2012, 4, 170-8.
[22] J Otukei & T Blaschke. Land Cover Change Assessment Using Decision Trees, Support
Vector Machines and Maximum Likelihood Classification Algorithms. Int. J. Appl. Earth
Observation. Geoformation. 2010, 12, 527-31.
[23] B Deilmai B Ahmad & H Zabihi. Comparison of Two Classification Methods (MLC & SVM)
to Extract Land Use and Land Cover in Johor city, Malaysia. IOP Conf. Ser. Earth Environ. Sci. 2014,
20, 1-6.
[24] H Dibs, M O Idrees, and G B A Alsalhin. 2017a Hierarchical Classification Approach for
Mapping Rubber Tree Growth Using Per-Pixel and Object-Oriented Classifiers with SPOT-5 Imagery
Egyptian Journal of Remote Sensing and Space Science 20 (1) 21–30.
https://doi.org/10.1016/j.ejrs.2017.01.004.
[25] H Dibs, A S Al-Janabi, C Gomes. 2017b. "Easy To Use Remote Sensing and GIS Analysis for
Landslide Risk Assessment." Journal of University of Babylon 26 (1): 42 - 54.

12
The Fourth Postgraduate Engineering Conference IOP Publishing
IOP Conf. Series: Materials Science and Engineering 745 (2020) 012166 doi:10.1088/1757-899X/745/1/012166

[26] H Dibs, M O Idrees, V Saeidi and S Mansor. 2016. "Automatic Keypoints Extractionfrom UAV
Image with Refine and Improved Scale Invariant Features Transform (RI-SIFT)." International Journal
of Geoinformatics 12 (3): 51-58.
[27] H Dibs, S Mansor, N Ahmed and B Pradhan 2015 Band-to-band registration model for near-
equatorial Earth observation satellite images with the use of automatic control point extraction Int J
Remote Sens 36 (8) 2184–2200.

13