View
22
Download
0
Category
Preview:
Citation preview
2017
Land Cover Classification Using Satellite Imagery and LiDAR
JONAS PUZINAS
Study program and semester: Masters of Geoinformatics 4th Semester
Project title:
Land cover classification using satellite imagery and LiDAR
Project period: 6th of February – 9th of June 2017
Supervisor(s): Jamal Jokar Arsanjani
Student: Jonas Puzinas
Number of copies:
Number of pages: 50
Abstract:
Remote sensing is commonly applied in land
cover mapping. Such maps can be created
using supervised or unsupervised classification
techniques onto digital images acquired by
remote sensing instruments, such as satellites.
During this project, the concept of introducing
the elevation information into the band
composite was tested in order to examine how
such ancillary data affects the accuracy of
classification methods.
Accuracy assessments using Confusion
Matrices were carried out and the results
provided that Support Vector Machine
classification perform better that the rest of
the tested classification techniques.
Aalborg University Copenhagen
A.C. Meyers Vænge 15
2450 Copenhagen SV
Secretary:
Tel:
E-mail:
Table of Contents
List of Figures .................................................................................................................................. 1
List of Tables ................................................................................................................................... 2
Acknowledgements ......................................................................................................................... 3
1. Introduction ............................................................................................................................ 4
2. Methodology ........................................................................................................................... 6
2.1. Software ........................................................................................................................... 7
2.2. Study Area ........................................................................................................................ 7
2.3. Data Collection ................................................................................................................. 8
2.3.1. Satellite Imagery ....................................................................................................... 8
2.3.2. LIDAR data ............................................................................................................... 11
2.4. Pre-processing ................................................................................................................ 13
2.4.1. Generation of Digital Surface Model ...................................................................... 14
2.4.2. Combining Bands .................................................................................................... 17
2.4.3. Training Samples ..................................................................................................... 18
2.5. Classification ................................................................................................................... 21
2.5.1. ISO Clustering .......................................................................................................... 21
2.5.2. Maximum Likelihood Classification ........................................................................ 21
2.5.3. Support Vector Machine Classification ................................................................... 21
2.6. Reclassification ............................................................................................................... 22
2.7. Confusion Matrix ............................................................................................................ 23
3. Results and Discussion .......................................................................................................... 27
3.1. Maps of Classification Outputs ...................................................................................... 27
3.2. Classification Accuracy Assessment ............................................................................... 34
3.2.1. ISO Cluster Unsupervised Classification ................................................................. 34
3.2.2. Maximum Likelihood Classification ........................................................................ 38
3.2.3. Support Vector Machine Classification ................................................................... 40
4. Conclusion ............................................................................................................................. 42
5. References ............................................................................................................................ 44
1
List of Figures
Figure 1 - Methodology scheme applied to this study ................................................................... 6
Figure 2 - Drawn polygon in EarthExplorer and setting Sentinel-2 satellite as source .................. 8
Figure 3 - The selected Sentinel-2 image marked by yellow rectangle .......................................... 9
Figure 4 – Acquired Sentinel-2 imagery data set: a) original; b) clipped ..................................... 11
Figure 5 – Point cloud data set collection (Kortforsyningen, 2017) ............................................. 12
Figure 6 - Point cloud elevation representation ........................................................................... 14
Figure 7 - Study area: Satellite imagery and DSM ........................................................................ 16
Figure 8 – Reference image samples (left - RGB, right - NDVI) ..................................................... 17
Figure 9 - Training samples for MLC in ground truth imagery: 1) Deep water 2) Shallow water 3)
Uncultivated fields 4) Built-up 5)Cultivated fields 6) Shrubs 7) Forest lands 8) Clouds ............... 19
Figure 10 - Spreading sampling points for SVM Classification ..................................................... 20
Figure 11 – Query to eliminate unwanted pixels .......................................................................... 24
Figure 12 - Frequency table .......................................................................................................... 25
Figure 13 - ISO Clustering output using Band Composite only ..................................................... 28
Figure 14 - ISO Clustering output using band composite with integrated DSM .......................... 29
Figure 15 - MLC output using Band Composite only .................................................................... 30
Figure 16 - MLC output with integrated DSM ............................................................................... 31
Figure 17 - SVM classification output using Band Composite only .............................................. 32
Figure 18 - SVM classification output with integrated DSM......................................................... 33
2
List of Tables
Table 1 - Information of Sentinel-2 bands (OSGeo, 2017)............................................................ 10
Table 2 - Point cloud classes (Flatman, et al., 2016) ..................................................................... 13
Table 3 - Support Vector Machine training accuracy ................................................................... 22
Table 4 - New land cover class codes after Reclassification ......................................................... 23
Table 5- Rearranged Frequency table ........................................................................................... 25
Table 6 – ISO unsupervised classification accuracy using band composite only .......................... 35
Table 7 - ISO unsupervised classification accuracy with integrated DSM .................................... 36
Table 8 - ISO unsupervised classification accuracy using band composite only after initial raster
was reclassified ............................................................................................................................. 37
Table 9 - ISO unsupervised classification accuracy with integrated DSM after initial raster was
reclassified .................................................................................................................................... 37
Table 10 - MLC accuracy using band composite only ................................................................... 38
Table 11 - MLC accuracy with integrated DSM ............................................................................. 39
Table 12 - SVM classification accuracy using band composite only ............................................. 40
Table 13 - SVM classification accuracy with integrated DSM ....................................................... 40
3
Acknowledgements
I would like to thank to some people who have helped me during the period of writing this thesis.
First of all, I would like to thank the supervisor of my project Jamal Jokar Arsanjani. I am very
grateful for his assistance, for suggesting the topic, advising on ideas of how to improve the study
and myself as a student. He also taught me of how to learn, kept encouraging me throughout and
most importantly inspired me with his positivity.
I would also like to thank my family and friends for supporting me during this time, for
encouraging to stay positive and concentrate on this work.
Last but not least, I am thankful to my colleagues at Laserpas for support and sharing their
experiences of doing Masters degree. Special thanks to my Head of Operations Daiva Kemzuraite
for encouragement and allowing some flexibility at work, and Bogdan Ghius, for sharing ideas
and comments regarding the topic of this thesis.
4
1. Introduction
Remote sensing became an immense subject among the modern geographic analysis methods.
It is used to detect a variation of objects on the surface of the Earth by using computer
implemented processing techniques (Wang, Wan, Ye, & Lai, 2017). It acquires and interprets
spectral, spatial, and temporal variation of electromagnetic energy emitted by a targeted subject
(Ghassemian, 2016). The production of land cover maps using image classification is one of the
conventional applications of remotely sensed imagery (Foody, 2002; Arsanjani & Vaz, 2015). A
definition of land cover in the INSPIRE Directive (INSPIRE Thematic Working Group Land Cover,
2013) defines it as a physical and biological cover of the Earth’s surface including classes such as
agricultural areas, forests, (semi-)natural or built-up areas, as well as wetlands, water bodies.
Land covers indicate the relationship between socio-economic activities and environmental
changes therefore it is essential to constantly update knowledge about the current state of the
landscape (Rujoiu-Mare & Mihai, 2016).
One of the means for acquiring imagery of the surface is satellites. Because of the fast developing
technology, satellites are now also able to retrieve high resolution images. In June of 2015,
European Space Agency launched Sentinel-2A satellite as their established mission. The satellite
is capturing images of the surface at high revisit frequency with high resolution 13 spectral bands
in order to provide information and contribute to land observations, emergency response and
security services, forestry and agricultural practises as well as to assist in food management
(European Space Agency, 2017; Drusch, et al., 2012). By frequently producing high quality images
of the Earth, it promotes new possibilities for research on land covers of the surface.
Numerous techniques of classification for land cover mapping are used in the field of remote
sensing and they can be divided into two types: supervised and unsupervised classification
(Wang, Wan, Ye, & Lai, 2017). The result of unsupervised classification is mainly dependent on
the characteristics of the input data set, on the other hand supervised classification requires
extensive sampling data for thorough identification of land cover classes (Beaubien, Cihlar, &
Latifovic, 1999). Two of the most commonly used supervised classification techniques are
5
Maximum Likelihood and Supervised Vector Machine classifications (Alajlan, Bazi, Melgani, &
Yager, 2012; Perronnin, Sanchez, & Mensink, 2010; Tarabalka, Benediktsson, & Chanussot, 2009).
However, large dimensional data spaces caused by advanced digital sensors can produce
classification inaccuracies and inconsistencies for digital land cover mapping (Alajlan, Bazi,
Melgani, & Yager, 2012; Beaubien, Cihlar, & Latifovic, 1999). To improve the accuracy, ancillary
data such as elevation could be used combined with spectral data (Elumnoh & Shrestha, 2000;
Ricchetti, 2000). Such combination provided significant increases in overall accuracy.
The focus of this study was to test three most widely used classification techniques, unsupervised
ISO clustering, supervised Maximum Likelihood and Support Vector Machine classifications, for
land cover mapping of Sentinel-2 imagery data combined together with Digital Surface Model.
Research question: How does Lidar data affect accuracy image classification for land cover?
The problems explored in the thesis:
1. What are the most common accuracy assessment approaches?
2. What is the most accurate classification method?
3. What land cover types are the easiest and most difficult to classify?
6
2. Methodology
The literature provided various approaches for land cover classification. By generalising the
findings, the methodology structure of this study was established and presented in Figure 1.
Figure 1 - Methodology scheme applied to this study
7
The workflow was started with collection and pre-processing of the data sets required for this
study. Later, training samples and signature files were generated. Finally, total of six
classifications were performed and final results evaluated.
This chapter will describe the system of methods respected in this project, from data acquisition
to final classified outputs as well as the accuracy assessment outcome. It will also comment on
software packages adapted in the research on the selected study area.
2.1. Software
The following programs were used to complete this study:
ArcMap 10.3
Bentley’s Microstation
Microsoft Excel
All classifications were utilised with ArcMap 10.3 Image classification toolbar. ArcMap aided
during the creation of composites as well as training samples (Esri (1), 2014). Two out of three
classification tools used were included in the basic software package. It needs to be emphasised
that an older version of ArcMap was used hence there was a need to install an additional toolbox
for Support Vector Machine classification, SVM tools by A. Wehmann (Wehmann, 2013; Chang &
Lin, 2011). Moreover, part of Confusion Matrices production was done with the same ArcMap
software.
Bentley’s Microstation was chosen for the Digital Surface Model generation from Lidar point
cloud as the author of the thesis is the most familiar with the use of this software.
For final result assessment, Confusion Matrices were applied. Although they were partially
generated within ArcMap, they needed to be refined with the aid of Microsoft Excel.
2.2. Study Area
For this project the North-eastern part of Zealand island in Denmark was selected due to the
area’s varying nature and therefore numerous land cover types can be found which was a key
8
aspect for this study. As a reinforcement, the required data sets for accomplishing this research
were fully available in the national database. Finally, this thesis is part of the Masters program at
the University of Aalborg, one of Denmark’s leading research environments and a strong
impactful institution for the author.
2.3. Data Collection
The idea of this thesis is to use two types of inputs, satellite imagery and surface model, both
available from two different sources. This section will describe where from and how data sets
were collected.
2.3.1. Satellite Imagery
Satellite Sentinel-2 imagery data set was available from United States Geological Survey (USGS)
and its online database Earth Explorer (United States Geological Survey, 2017). Firstly, the
polygon representing the area of interest was drawn and Sentinel-2 satellite as the Source was
chosen (Figure 2).
Figure 2 - Drawn polygon in EarthExplorer and setting Sentinel-2 satellite as source
9
From suggested results, the data set with the least cloud cover was chosen which appeared to
cover the area as presented in Figure 3 below.
Figure 3 - The selected Sentinel-2 image marked by yellow rectangle
The used imagery was acquired on the 11th of March, 2017 by Sentinel-2 satellite of Copernicus
Space program (European Space Agency, 2017). Since March is the spring season in Denmark
(Bruun & Ejrnæs, 2000) and vegetation has started to grow, this data set was at sufficient quality
for vegetation to be distinguished in the images. Furthermore, it was possible to use imagery
data that was acquired in the middle or late summer but priority was given to data with the least
cloud cover. Nonetheless, the selected data set included 13 spectral bands with spatial resolution
of each varying between 10, 20 or 60 meters (Table 1), and less than 10% cloud cover. Each image
was provided in JPEG2000 format and WGS84 UTM 33N (EPSG: 326330) map projection system
(United States Geological Survey Metadata, 2017).
10
Table 1 - Information of Sentinel-2 bands (OSGeo, 2017)
Original size of the image retrieved from Earth Explorer included some part of Sweden therefore
it has been clipped in order to reduce it to the area of interest (Figure 4) which eventually covered
approximately 6690 square kilometres including both land and water.
11
a)
b)
Figure 4 – Acquired Sentinel-2 imagery data set: a) original; b) clipped
2.3.2. LIDAR data
The second data set, LIDAR point cloud, was collected from Danish government database
Kortforsyningen (Styrelsen for Dataforsyning og Effektivisering, 2017). As opposed to satellite
imagery, LIDAR point cloud area was selected only to cover a small part of the study area. It was
acquired in 2015 and grouped into a 10 by 10 kilometre block covering an area of Gentofte (Figure
5), georeferenced to WGS 84 UTM32N projection system and provided in LAZ format.
12
Figure 5 – Point cloud data set collection (Kortforsyningen, 2017)
The point cloud’s average density was 4.5 points per square meter, with vertical accuracy of 5 cm
and horizontal accuracy 15 cm (Miljøministeriet, Geodatastyrelsen, 2015). The point cloud was
within the area of interest, 105.5 square kilometres in size which is 1.6% of the area of interest
and it included the following codification for the classes (Table 2):
13
Table 2 - Point cloud classes (Flatman, et al., 2016)
2.4. Pre-processing
Data pre-processing is needed to transform the raw unorganized information into an
understandable format for further analysis. The preliminary section of the project included: 1)
Lidar point cloud conversion to Digital Surface Model; 2) creation of composite bands out of
acquired satellite imagery data set and DSM. This section will outline the completion of
aforementioned processes prior to the actual conducted classifications.
14
2.4.1. Generation of Digital Surface Model
In order to be able to integrate height information, Lidar point cloud had to be converted into
DSM (Chen, Su, Li, & Sun, 2009; Huang, Zhang, & Gong, 2011). Firstly, the point cloud was loaded
into Bentley’s Microstation environment (Figure 6) and re-projected to WGS84 UTM 33N to
match satellite imagery’s spatial reference.
Figure 6 - Point cloud elevation representation
15
Secondly, for exporting purposes, the following classes were selected in order to best represent
area’s surface elevation: 2) Ground; 3) Low vegetation; 4) Medium vegetation; 5) High
vegetation; 6) Buildings and structures; 9) Water; 17) Bridge decks. Finally, DSM raster was
exported with 10 meter cell size and an individual cell representing height value (Figure 7).
16
Figure 7 - Study area: Satellite imagery and DSM
17
2.4.2. Combining Bands
The total of four composites used for this study were produced using ArcMap 10.3:
1) Composite of Sentinel-2 bands 2, 3, 4 (RGB)
2) Composite of Sentinel-2 bands 3, 4, 8 (Normalised Difference Vegetation Index)
3) Composite of Sentinel-2 bands 2-8, 8A, 11-12
4) Composite of Sentinel-2 bands 2-8, 8A, 11-12 and DSM
1) and 2) composites were used as ground truth data sets in creation of training samples and
accuracy assessment. While 1) was used as true colour to inspect actual locations within training
samples, 2) was useful to recognise different land covers that included vegetation.
Figure 8 – Reference image samples (left - RGB, right - NDVI)
3) and 4) were used as input rasters for different classification methods to be tested. To create
these composites, bands 2-8, 8A, 11-12 were combined because bands 1, 9 and 10 (refer to Table
1 in Section 2.3.1.) did not provide height information and their spatial resolution of 60 meters
was too low (Sentinel, ESA, 2017; Congedo, 2016). When compiling the second testing composite,
DSM was inserted as a spectral band.
18
2.4.3. Training Samples
There were two different types of training samples collected for the use of supervised
classifications. Both sets were represented by vectorial layers, first – by polygons, the other – by
points. Training polygons were used for Maximum Likelihood Classification. They were created
using standard procedure through ArcMap 10.3 Image Classification tool by drawing them over
the areas of each of distinguished land cover (Esri (2), 2014) as shown in Figure 9. In order to
locate and identify each land cover class properly, RGB and NDVI rasters facilitated the creation
of training samples. Later, the training samples were used to generate signature files for both
supervised classifications.
19
1)
2)
3)
4)
5)
6)
7)
8)
Figure 9 - Training samples for MLC in ground truth imagery: 1) Deep water 2) Shallow water 3) Uncultivated fields 4) Built-up
5)Cultivated fields 6) Shrubs 7) Forest lands 8) Clouds
20
Additionally, individual point samples were collected for Support Vector Machine classification.
By following instructions of SVM Tools for ArcGIS 10.1+ (Wehmann, 2013), ArcMap Create
Random Points tool was utilised to spread 1000 points within imagery area boundaries (Figure
10).
Figure 10 - Spreading sampling points for SVM Classification
21
Then, same as for training polygons, class values 1-8 were manually assigned for each point
depending on which land cover a point lied to the reference of RGB and NDVI composites
mentioned previously. Subsequently, the points were converted to the training raster thus
creating pixels instead of points which held land cover code as attribute.
2.5. Classification
2.5.1. ISO Clustering
ISO Clustering is unsupervised classification therefore there was minimum requirement for user’s
intervention. It was necessary to select an input raster and a number of classes into which cells
will be grouped. After these parameters were set, the tool performed the classification and the
classified output raster was retrieved (Esri (3), 2014). However, for the reasons stated in Section
3.2.1., post-processing was required which will be described in Section 2.6.
2.5.2. Maximum Likelihood Classification
This supervised classification method required signature file as an input. Two signature files were
generated by loading already created training sample polygons (Section 2.4.3.) on each of the
testing composite bands consisting of 10 Sentinel bands and an addition of DSM (Section 2.4.1.)
(Esri (4), 2014). When utilising the tool, it then assigned each cell of the composites to the classes
represented in the signature files proportionally to the number of cells apprehended by each
signature (Esri (5), 2014).
2.5.3. Support Vector Machine Classification
SVM tool was applied following two phase procedure: 1) running the tool without generating
output raster therefore SVM is only trained, 2) after the accuracy no longer increases, the tool is
executed to produce classified raster (Wehmann, 2013). While MLC required training samples in
order to generate signature files, alternatively SVM used sampling points to generate output
models after each training. The tool provided the possibility to train SVM until it achieved the
best possible accuracy thus three stages of training were conducted (Table 3).
22
Training Band Composite only With infused DSM
1 81,7% 81,9%
2 81.8% 82,5%
3 81.8% 82,5%
Table 3 - Support Vector Machine training accuracy
The computed results showed that SVM improved the accuracy after two trainings and the result
of last training indicated that it will no longer developed and reached its peaks.
Finally, in order to obtain SVM classification output, the output model of the second training was
used as well as the composites and the training rasters.
2.6. Reclassification
This step was only applied to ISO Clustering during all the process. When evaluating the results
(will be provided later in Section 3.2.1) of unsupervised classification, there was a need to
reassign values of the classified raster. It was visually inspected that the class codes defined by
the user did not match the ones produced by this classification. Firstly, the amount of classes was
reduced to 7 by eliminating Cloud class as it was unstable and blended into other land cover
classes. Then, using ArcMap 10.3 Reclassify tool new pixel values were designated. While Deep
water, Shallow water and Built-up classes were left the same, the rest of the newly assigned
classes were as follow in the Table 4 below.
23
Initial classes Re-assigned classes
3) Uncultivated fields 7) Forest lands
5) Cultivated fields 6) Shrubs
6) Shrubs 3) Uncultivated fields
7) Forest lands 3) Uncultivated fields
8) Clouds 5) Cultivated fields
Table 4 - New land cover class codes after Reclassification
2.7. Confusion Matrix
In this study, Confusion Matrices were used to assess classification accuracy of remotely sensed
data. Outputs of the three tested classification methods were compared against ground truth
data. To accomplish this, ArcMap 10.3 and Microsoft Excel were used (Verbyla, 2013).
Firstly, the values of classified rasters were extracted using ArcMap 10.3 Extract Values to Points
tool. For the locations where the values were compared, Training Sample file which was created
for SVM sampling and containing point features was used as the input. The tool produced a
feature class that contained both ground truth and predicted values for the same location.
However, there were some pixels that had to be eliminated because they had no data values
therefore it returned -9999 in the attribute table. The elimination was carried out by simply
building a query displayed in Figure 11, which resulted in disposing of 5 pixels out of total of 1477.
24
Figure 11 – Query to eliminate unwanted pixels
The second step was to summarise how many predictions were made for every ground truth
class. To do this, Frequency table was used and the result is displayed in Figure 12.
25
Figure 12 - Frequency table
Furthermore, the last step in ArcMap was to transform this information into a Confusion Matrix
arrangement using Pivot table which re-organised the results shown in Table 5.
Table 5- Rearranged Frequency table
26
Lastly, in order to finalise the tables, they were transported into Microsoft Excel, names of
columns and rows were changed and presented in Results Chapter.
27
3. Results and Discussion
This chapter will display the final land cover maps, generated by the three classification methods
in scope, that resulted in applying established methodology. Furthermore, a discussion on
accuracy assessment through application of Confusion matrices will follow.
3.1. Maps of Classification Outputs
This subsection presents maps containing final outputs, i.e. classified rasters, that were
generated after following the methodology described in Chapter 3.
28
Figure 13 - ISO Clustering output using Band Composite only
29
Figure 14 - ISO Clustering output using band composite with integrated DSM
30
Figure 15 - MLC output using Band Composite only
31
Figure 16 - MLC output with integrated DSM
32
Figure 17 - SVM classification output using Band Composite only
33
Figure 18 - SVM classification output with integrated DSM
34
3.2. Classification Accuracy Assessment
After all the classified rasters were gathered, accuracy assessment was carried out using
Confusion Matrix method. Excel tables that assisted for calculations will be presented in this
section as well as discussion on the results.
The first column from the left presents the classes the classification resulted in, while the top row
presents the ground truth data.
Main focus should be given to the values in blue fields in tables presented, that provide total
number of accurately classified pixels of the image from randomly collected samples. The cell in
yellow shows the overall accuracy classification. The criteria for this value is 85% (Thomlinson,
Bolstad, & Cohen, 1999). The first column on the right displays Producer’s accuracies that
represent the accuracy of each class being classified by the classifier. If it is 75% for land cover
type A (Thomlinson, Bolstad, & Cohen, 1999), it means that as a producer of data, I can trust my
classification for that particular land cover type. Also, the last row, which presents User’s
accuracies, estimates the reliability of the classification for each land type for a user of the
classification result.
3.2.1. ISO Cluster Unsupervised Classification
The resulting classification outputs are presented in Section 3.1 of this Chapter in Figures 13 and
14, and Confusion Matrix tables 6 and 7. The tables depict Unsupervised ISO classification
accuracy when inputs are: 1) 13-band composite raster 2) the same raster with additionally
infused DSM that covers just a small part of the overall area. Since this classification method is
not influenced by a user, it determines groups of clusters, not actual classes, and is dependent
solely on image raster provided by the user. For this reason, apart from the number of classes, it
is difficult to foresee what output this classification will generate.
35
Table 6 – ISO unsupervised classification accuracy using band composite only
The results indicate a fluctuation in producer’s and user’s accuracy. Producer’s accuracy show
that part of land covered by water (classes Deep water and Shallow water) were recognised most
accurately among all, 54.8% and 53.4%, then followed by urban Built-up and Uncultivated fields
classes which turned out to be 27.8% and 25% accordingly. Clouds were correctly recognised in
just above 22 percent of the samples. However, the method showed no precision on detecting
Forest lands correctly while only 2.7% of cells presenting Cultivated fields were correctly
classified.
On the other hand, user’s accuracy is only high for Deep water class (98%), which is then followed
by 30.1% for Shallow water, 20.4% for Clouds, and unfortunately drops to 0% for Forest lands,
same as seen in producer’s accuracy.
Confusion matrix reveals what confusions arose for ISO classification when distinguishing classes.
For instance, 145 pixels out of 356 were considered to be Deep water when they actually were
in the class of Shallow water. On the other hand, relatively large part of the Shallow water was
considered to be Clouds. Unfortunately, ISO method faced even bigger problems while predicting
the rest of the land covers. Built-up areas were greatly mixed up with Uncultivated fields and
some part of Clouds as well as Shrub class was largely confused with Built-up class, and Cloud
class with Cultivated fields. Finally, Forest lands were not correctly predicted at all and this class
was mainly baffled by Cultivated fields.
As it is seen, the unsupervised cluster classification showed no reliability in correctly predicting
classes and the total accuracy of 24.3% proves it.
Class NameDeep
water
Shallow
water
Uncultivated
fieldsBuilt-up
Cultivated
fieldsShrubs
Forest
landsClouds
Producer's
accuracyPredictions
1) Deep water 195 145 0 0 0 0 0 16 54,8% 356
2) Shallow water 4 63 1 1 1 0 6 42 53,4% 118
3) Uncultivated fields 0 1 14 17 0 15 5 4 25,0% 56
4) Built-up 0 0 54 30 1 1 0 22 27,8% 108
5) Cultivated fields 0 0 31 69 11 144 157 0 2,7% 412
6) Shrubs 0 0 43 50 15 16 0 10 11,9% 134
7) Forest lands 0 0 47 12 83 5 0 15 0,0% 162
8) Clouds 0 0 8 11 74 2 3 28 22,2% 126
User's accuracy 98,0% 30,1% 7,1% 15,8% 5,9% 8,7% 0,0% 20,4% 24,3% total accuracy
Count Truth 199 209 198 190 185 183 171 137 1472
36
However, when classifying the same image with infused Digital Surface Model, overall accuracy
has been increased only by 0.2% (Table 7). The error matrix table show that the predictions by
the producer were slightly improved on Shrubs (from 11.9% to 14.3%) and Cultivated fields (from
2.7% to 4.4%) but the rest of the classification proved even less accuracy on predicting all other
land covers, emphasising that Forest lands were still atrociously imprecise with the same 0%
accuracy.
Unfortunately, user’s accuracy also showed no significant development after addition of the
DSM. It only increased for Cultivated fields, Shrub and slightly for Clouds classes. Despite that,
the accuracy decreased for both water body classes, and stayed at zero percent for the same
Forest land class.
Table 7 - ISO unsupervised classification accuracy with integrated DSM
Nonetheless the accuracy of the unsupervised classification until now, a suspicion appeared, that
the algorithm could use different cluster number assignment and that would explain low
percentage in predicting pixel values. To investigate this, reclassification was done, as described
in Section 2.6, in order to artificially bring classified raster’s values closer to the reference data
set. After this step was conducted, new confusion matrices were generated (Tables 8 and 9) and
the accuracy was re-examined.
Class NameDeep
water
Shallow
water
Uncultivated
fieldsBuilt-up
Cultivated
fieldsShrubs
Forest
landsClouds
Producer's
accuracyPredictions
1) Deep water 194 151 0 0 0 0 0 16 53,7% 361
2) Shallow water 3 56 1 1 1 0 6 44 50,0% 112
3) Uncultivated fields 2 2 14 17 0 17 5 2 23,7% 59
4) Built-up 0 0 54 30 1 2 0 22 27,5% 109
5) Cultivated fields 0 0 33 66 18 136 156 0 4,4% 409
6) Shrubs 0 0 43 51 16 20 0 10 14,3% 140
7) Forest lands 0 0 45 12 76 6 0 14 0,0% 153
8) Clouds 0 0 8 13 73 2 4 29 22,5% 129
User's accuracy 97,5% 26,8% 7,1% 15,8% 9,7% 10,9% 0,0% 21,2% 24,5% total accuracy
Count Truth 199 209 198 190 185 183 171 137 1472
37
Table 8 - ISO unsupervised classification accuracy using band composite only after initial raster was reclassified
Table 9 - ISO unsupervised classification accuracy with integrated DSM after initial raster was reclassified
Re-assigning pixel values has greatly increased total accuracy as well as user’s and producer’s
when classifying both, composite of bands only and with added DSM. When previously total
accuracy values were 24.3% and 24.5%, now they are above increased to 45% and 43.6%. While
percentage of predicted values for Deep water class was at the same level, predictions for the
rest of the classes were extremely improved. However, after reclassification of initially classified
raster, integration of DSM did not develop better accuracy. Contrary to first classification, it
reduced from 45% to 43.6%. This reveals, that it is complicated to evaluate the accuracy of
unsupervised classification using Confusion Matrix because it requires additional steps to purify
the output.
All in all, covers such as Forest or Built-up lands, that contain objects which are higher than usual
vegetation like in Cultivated or Uncultivated fields, were not predicted notably better. It was
expected that DSM generated from Lidar point cloud should highlight the difference between
classes and benefit the algorithm during classification. However, the results suggest that overall,
Class NameDeep
water
Shallow
water
Uncultivated
fieldsBuilt-up
Cultivated
fieldsShrubs
Forest
lands
Producer's
accuracyPredictions
1) Deep water 195 145 0 0 0 0 0 57,4% 340
2) Shallow water 4 63 1 1 1 0 6 82,9% 76
3) Uncultivated fields 0 0 90 62 98 21 0 33,2% 271
4) Built-up 0 0 54 30 1 1 0 34,9% 86
5) Cultivated fields 0 0 8 11 74 2 3 75,5% 98
6) Shrubs 0 0 31 69 11 144 157 35,0% 412
7) Forest lands 0 1 14 17 0 15 5 9,6% 52
User's accuracy 98,0% 30,1% 45,5% 15,8% 40,0% 78,7% 2,9% 45,0% total accuracy
Count Truth 199 209 198 190 185 183 171 1335
Class NameDeep
water
Shallow
water
Uncultivated
fieldsBuilt-up
Cultivated
fieldsShrubs
Forest
lands
Producer's
accuracyPredictions
1) Deep water 194 151 0 0 0 0 0 56,2% 345
2) Shallow water 3 56 1 1 1 0 6 82,4% 68
3) Uncultivated fields 0 0 88 63 92 26 0 32,7% 269
4) Built-up 0 0 54 30 1 2 0 34,5% 87
5) Cultivated fields 0 0 8 13 73 2 4 73,0% 100
6) Shrubs 0 0 33 66 18 136 156 33,3% 409
7) Forest lands 2 2 14 17 0 17 5 8,8% 57
User's accuracy 97,5% 26,8% 44,4% 15,8% 39,5% 74,3% 2,9% 43,6% total accuracy
Count Truth 199 209 198 190 185 183 171 1335
38
DSM had no influence on unsupervised classification while it was anticipated that such ancillary
data should improve the results.
3.2.2. Maximum Likelihood Classification
Unlike the previously discussed unsupervised classification, confusion matrix of Maximum
Likelihood Classification (Table 10) illustrates a different scenario. The highest accuracy using the
composite of bands was achieved for Forest lands and Cultivated fields, accordingly 99.3% and
92.1%. Surprisingly, Shrub class was well distinguished from the two aforementioned classes
even when they are visually similar. While both classes of water bodies, Uncultivated fields and
Clouds were predicted almost the same, accuracy for Built-up areas barely reached 54%. It could
be explained by variety of different spectral characteristics within the class.
However, there was an analogy between ISO clustering and MLC when some pixels of Deep water
were mixed up with Shallow water (49 pixels). Similarly, 70 pixels were considered to be Built-up
land where the ground truth was Uncultivated fields.
Comparing Maximum Likelihood Classification with ISO clustering using only band composite as
an input raster, the MLC producer’s accuracy was a lot more accurate with the total value of
76.8%.
Moreover, user’s accuracy was much closer to the satisfactory level than for ISO clustering. In
more detail, Uncultivated fields and Shallow water were the least reliable classes from this point
of view, while other classes showed much higher, 70% and above, reliability with Shrubs reaching
91.3%.
Table 10 - MLC accuracy using band composite only
Class NameDeep
water
Shallow
water
Uncultivated
fieldsBuilt-up
Cultivated
fieldsShrubs
Forest
landsClouds
Producer's
accuracyPredictions
1) Deep water 170 49 0 0 0 0 0 1 77,3% 220
2) Shallow water 27 127 0 0 0 0 0 11 77,0% 165
3) Uncultivated fields 0 0 104 6 29 3 1 1 72,2% 144
4) Built-up 0 14 70 175 26 12 11 17 53,8% 325
5) Cultivated fields 0 0 8 1 129 1 0 1 92,1% 140
6) Shrubs 0 0 16 6 0 167 2 0 87,4% 191
7) Forest lands 0 0 0 0 1 0 152 0 99,3% 153
8) Clouds 2 19 0 2 0 0 5 106 79,1% 134
User's accuracy 85,4% 60,8% 52,5% 92,1% 69,7% 91,3% 88,9% 77,4% 76,8% total accuracy
Count Truth 199 209 198 190 185 183 171 137 1472
39
When using another image raster with infused Digital Surface Model, confusion matrix below in
Figure 11 showed overall producer’s accuracy of 76.8%. There were three land cover classes
predicted at accuracies above 90%: Cultivated fields (91%), Forest lands (98.7%) and Shallow
water (100%). As before, when input raster did not include Digital Surface Model, one of the two
poorly predicted classes was Built-up. The latter along with Cloud class only reached 50.4% and
39.9% accordingly. This again proves that Built-up land cover is the most mixed in terms of
spectral reflectance.
Table 11 - MLC accuracy with integrated DSM
On the contrary, user’s accuracy for each class has changed comparing to the one with classified
composite of bands only. On aggregate, they all dropped and overall user’s accuracy went down
from 77.4%, when input was band composite only, to 69.6%, when input raster additionally
contained DSM.
However, considering that Maximum Likelihood Classification training samples consisted of
polygons and not individual pixels, this allows higher probability for the user to include pixels of
different land cover and therefore incorrectly defining a reference land cover class.
Furthermore, the MLC results show that Digital Surface Model did not improve the classification
accuracy. Instead, it decreased the classifier’s quality in detecting land types properly.
To sum up the results of the two already discussed classification methods, unsupervised ISO and
supervised MLC, there are two main aspects that give such disparity. Firstly, ISO clustering
considers statistical properties of the pixels and groups them into clusters. Secondly, Maximum
Likelihood Classification takes into account training samples from user which provide information
about pixels, therefore it greatly improves the accuracy of classification.
Class NameDeep
water
Shallow
water
Uncultivated
fieldsBuilt-up
Cultivated
fieldsShrubs
Forest
landsClouds
Producer's
accuracyPredictions
1) Deep water 179 67 0 0 0 0 0 3 71,9% 249
2) Shallow water 0 25 0 0 0 0 0 0 100,0% 25
3) Uncultivated fields 0 0 97 4 31 4 1 1 70,3% 138
4) Built-up 2 9 63 135 29 16 9 5 50,4% 268
5) Cultivated fields 0 0 9 1 122 1 0 1 91,0% 134
6) Shrubs 0 0 19 5 0 159 2 0 85,9% 185
7) Forest lands 0 0 0 0 2 0 153 0 98,7% 155
8) Clouds 18 108 10 45 1 3 6 127 39,9% 318
User's accuracy 89,9% 12,0% 49,0% 71,1% 65,9% 86,9% 89,5% 92,7% 67,7% total accuracy
Count Truth 199 209 198 190 185 183 171 137 1472
40
3.2.3. Support Vector Machine Classification
As seen in the tables 12 and 13 below, the accuracy ranges for either input rasters were above
84% where none of the previously reviewed classification methods fell within.
Table 12 - SVM classification accuracy using band composite only
The total accuracy was 84.2% when band composite only was used as an input raster. However,
opposite to other classification results where Deep water classes had the highest accuracy, this
time it was one of the least predicted classes. Together with Uncultivated fields, producer’s
accuracy did not reach 75%. Nonetheless, all the rest of the predictions where in between 84.9%
for Shallow water to 95% for Forest lands.
Examining user’s accuracy with the same input, it was only of Shallow water with the accuracy of
72.7% that had the least reliability. The rest of the classes were evaluated above 75%. The firmest
ground truths were for Deep water, Shrubs and Forest lands as they hit above 88%.
Table 13 - SVM classification accuracy with integrated DSM
When input was the raster with incorporated DSM, the total producer’s and user’s accuracy has
increased by 0.4% and 0.5% accordingly. DSM helped the producer to better predict pixels of
Class NameDeep
waterShallow water
Uncultivated
fieldsBuilt-up
Cultivated
fieldsShrubs
Forest
landsClouds
Producer's
accuracyPredictions
1) Deep water 182 55 0 0 0 0 0 6 74,9% 243
2) Shallow water 17 152 0 0 0 0 3 7 84,9% 179
3) Uncultivated fields 0 1 154 24 15 15 3 1 72,3% 213
4) Built-up 0 0 16 143 6 0 0 3 85,1% 168
5) Cultivated fields 0 0 11 8 163 2 0 2 87,6% 186
6) Shrubs 0 0 13 6 0 165 2 0 88,7% 186
7) Forest lands 0 1 2 3 1 1 163 0 95,3% 171
8) Clouds 0 0 2 6 0 0 0 118 93,7% 126
User's accuracy 91,5% 72,7% 77,8% 75,3% 88,1% 90,2% 95,3% 86,1% 84,2% total accuracy
Count Truth 199 209 198 190 185 183 171 137 1472
Class NameDeep
water
Shallow
water
Uncultivated
fieldsBuilt-up
Cultivated
fieldsShrubs
Forest
landsClouds
Producer's
accuracyPredictions
1) Deep water 182 55 0 0 0 0 0 6 74,9% 243
2) Shallow water 17 152 0 0 0 0 4 6 84,9% 179
3) Uncultivated fields 0 1 149 20 15 11 3 1 74,5% 200
4) Built-up 0 0 16 150 6 0 0 3 85,7% 175
5) Cultivated fields 0 0 12 9 163 2 0 2 86,7% 188
6) Shrubs 0 0 18 5 0 168 1 0 87,5% 192
7) Forest lands 0 1 1 4 1 2 163 0 94,8% 172
8) Clouds 0 0 2 2 0 0 0 119 96,7% 123
User's accuracy 91,5% 72,7% 75,3% 78,9% 88,1% 91,8% 95,3% 86,9% 84,6% total accuracy
Count Truth 199 209 198 190 185 183 171 137 1472
41
Uncultivated and Built-up land covers as well as Clouds. However, the producer’s accuracy
decreased for Cultivated fields, Shrubs and Forest lands.
Alternatively, DSM notably affected reliability of three classes: Uncultivated fields, Built-up and
Shrubs. The biggest increase in user’s accuracy was for Built-up land cover, where it went up from
75.3% to 78.9%. Then it is followed by Shrub cover, where it developed by 1.6%. However,
introduction of DSM negatively affected the user’s accuracy for Uncultivated fields – it decreased
by the largest increment of 2.5%.
To evaluate the effect of presence of height information in classified raster, it did influence the
total accuracy of Support Vector Machine classification method. Both, user’s and producer’s
overall accuracy has increased. Confusion matrices of SVM classification accuracy gave the best
results among all classification methods used for this project, and it was affirmed that Support
Vector Machine classification was the most advanced method.
42
4. Conclusion
In this report the idea of remote sensing was presented. It acquires and interprets spectral and
spatial information reflected by the surface of the Earth and it is commonly applied in land cover
mapping. Land covers are physical and biological covers of the Earth’s surface and they include
natural, (semi-) natural or human-built areas. Land covers are also indicators of the relationship
between environmental changes and socio-economic activities. The report also introduced to
Sentinel-2 satellite and its importance in advancing the idea of land cover mapping.
Furthermore, unsupervised and supervised digital image classifications and their differences
were presented. Additionally, a possible solution of their accuracy was discussed and the
methodology for this study was established.
The concept of introducing the elevation information into the band composite was tested.
However, by following the methodology established for this project and encountering software
limitations, anticipated results were not achieved. The accuracy assessment for 2 out of 4 tests
provided minor improvements.
Support Vector Machine classification showed the trend that addition of Digital Surface Model
combined along with colour bands retrieved from Sentinel 2 satellite improves the accuracy of
classification. Respecting that infused DSM was only 1.6% size of the total study area, the increase
in accuracy of 0.4% proposes that this technique was applied effectively.
Unsupervised ISO Clustering provided the trend of improvement in accuracy during initial
procedure. On the other hand, after the reclassification the pattern was breached. Confusion
matrix suggested that the established method for this study was not sufficient and required
purification of classification outcomes. For example, a more thorough inspection of initial
classification output before reassigning classes.
The hypothesis again was denied when assessing accuracy of Maximum Likelihood Classification
output. The accuracy notably dropped when DSM was introduced to the raster composite. To the
respect that MCL classification is advocated to be reliable method in the literature, the results of
this study proposed a perception that the user misconducted the technique. As the only input
43
from user side was the signature file, thus training samples could be improved by collecting
smaller samples and consequently reducing the error possibly caused by including pixels of other
land covers.
44
5. References
Alajlan, N., Bazi, Y., Melgani, F., & Yager, R. R. (2012). Fusion of supervised and unsupervised
learning for improved classification of hyperspectral images. Information Sciences, 217,
39-55. doi:http://dx.doi.org/10.1016/j.ins.2012.06.031
Arsanjani, J. J., & Vaz, E. (2015). An assessment of a collaborative mapping approach for exploring
land use patterns for several European metropolises. International Journal of Applied
Earth Observation and Geoinformation, 35, 329-337.
Beaubien, J., Cihlar, J., & Latifovic, R. (1999, November 1). Land cover from multiple thematic
mapper scenes using a new enhancement-classification methodology. Journal of
Geographical Research, 104(D22), 27909-27920. doi:10.1029/1999JD900243
Bruun, H. H., & Ejrnæs, R. (2000, August). Classification of dry grassland vegetation in Denmark.
Journal of Vegetation Science, 11(4), 585-596. doi:10.2307/3246588
Chang, C. C., & Lin, C. J. (2011). LIBSVM: a library for support vector machines. ACM Transactions
on Intelligent Systems and Technology, 2(27), 1-27. Retrieved from
http://www.csie.ntu.edu.tw/~cjlin/libsvm
Chen, Y., Su, W., Li, J., & Sun, Z. (2009, April 1). Hierarchical object oriented classification using
very high resolution imagery and LIDAR data over urban areas. Advances in Space
Research, 43(7), 1101-1110.
Congalton, R. G. (1991, July). A review of assessing the accuracy of classifications of remotely
sensed data. Remote Sensing of Environment, 37(1), 35-46.
Congedo, L. (2016). Semi-Automatic Classification Plugin Documentation.
doi:http://dx.doi.org/10.13140/RG.2.2.29474.02242/1
Drusch, M., Del Bello , U., Carlier, S., Colin, O., Fernandez, V., Gascon, F., . . . Bargellini, P. (2012).
Sentinel-2: ESA's Optical High-Resolution Mission for GMES Operational Services. Remote
Sensing of Environment, 120, 25-36. doi:10.1016/j.rse.2011.11.026
45
Elumnoh, A., & Shrestha, R. P. (2000, March). Application of DEM Data to Landsat Image
Classification: Evaluation in a Tropical Wet-Dry Landscape of Thailand. Photogrammetric
Engineering & Remote Sensing, 66(3), 297-304. doi:0099-1112/00/6503-297$3.00/0
Esri (1). (2014). Image classification using the ArcGIS Spatial Analyst extension, 10.3. Retrieved
2017, from http://desktop.arcgis.com:
http://desktop.arcgis.com/en/arcmap/10.3/guide-books/extensions/spatial-
analyst/image-classification/image-classification-using-spatial-analyst.htm
Esri (2). (2014). Creating training samples, 10.3. Retrieved 2017, from http://desktop.arcgis.com:
http://desktop.arcgis.com/en/arcmap/10.3/guide-books/extensions/spatial-
analyst/image-classification/creating-training-samples.htm
Esri (3). (2014). Iso Cluster Unsupervised Classification, 10.3. (Esri) Retrieved 2017, from
http://desktop.arcgis.com/: http://desktop.arcgis.com/en/arcmap/10.3/tools/spatial-
analyst-toolbox/iso-cluster-unsupervised-classification.htm
Esri (4). (2014). Creating a signature file, 10.3. Retrieved 2017, from http://desktop.arcgis.com/:
http://desktop.arcgis.com/en/arcmap/10.3/guide-books/extensions/spatial-
analyst/image-classification/creating-training-samples.htm
Esri (5). (2014). How Maximum Likelihood Classification works. Retrieved 2017, from
http://desktop.arcgis.com/: http://desktop.arcgis.com/en/arcmap/10.3/tools/spatial-
analyst-toolbox/how-maximum-likelihood-classification-works.htm
European Space Agency. (2017). Retrieved from http://www.esa.int:
http://www.esa.int/Our_Activities/Observing_the_Earth/Copernicus/Sentinel-
2/Introducing_Sentinel-2
Flatman, A., Rosenkranz, B., Evers, K., Bartels, P., Kokkendorff, S., Knudsen, T., & Nielsen, T.
(2016). Quality assessment report to the Danish Elevation Model (DK-DEM). Quality
assessment report, Agency for Data Supply and Efficiency. Retrieved from
http://kortforsyningen.dk/indhold/data
46
Foody, G. M. (2002). Status of land cover classification accuracy assessment. Remote Sensing of
Environment, 80, 185-201.
Ghassemian, H. (2016). A review of remote sensing image fusion methods. Information Fusion,
32, 75-89. doi:http://dx.doi.org/10.1016/j.inffus.2016.03.003
Huang, X., Zhang, L., & Gong, W. (2011). Information fusion of aerial images and LIDAR data in
urban areas: vector-stacking, re-classification and post-processing approaches.
International Journal of Remote Sensing, 32(1), 69-84. doi:10.1080/01431160903439882
INSPIRE Thematic Working Group Land Cover. (2013). D2.8.II.2 INSPIRE Data Specification on
Land Cover – Technical Guidelines. European Commission Joint Research Centre.
Retrieved 2017, from
http://inspire.ec.europa.eu/documents/Data_Specifications/INSPIRE_DataSpecification
_LC_v3.0.pdf
Kortforsyningen. (2017, April). Retrieved from https://download.kortforsyningen.dk/:
https://download.kortforsyningen.dk/content/dhmpunktsky
Miljøministeriet, Geodatastyrelsen. (2015). Danmarks Højdemodel, DHM/Punktsky. Product
Specification, Miljøministeriet, Geodatastyrelsen, Copenhagen. Retrieved from
https://kortforsyningen.dk/sites/default/files/old_gst/DOKUMENTATION/Data/dk_dhm
_punktsky_v2_jan_2015.pdf
Muller, S. V., Walker, D. A., Nelson, F. E., Auerbach, N. A., Bockheim, J. G., Guyer, S., & Sherba, D.
(1998, June). Accuracy Assessment of a land-cover map of The Kuparuk River Basin,
Alaska: Considerations for remote regions. Photogrammetric Engineering and Remote
Sensing, 64(6), 619-628.
OSGeo. (2017). Retrieved April 2017, from www.gdal.org:
http://www.gdal.org/frmt_sentinel2.html
Perronnin, F., Sanchez, J., & Mensink, T. (2010). Improving the Fisher Kernel for Large-Scale Image
Classification. ECCV 2010 - European Conference on Computer Vision (pp. 143-156).
Heraklion, Greece: Springer-Verlag. doi:10.1007/978-3-642-15561-1 11
47
Ricchetti, E. (2000, April). Multispectral Satellite Image and Ancillary Data Integration for
Geological Classification. Photogrammetric Engineering & Remote Sensing, 66(4), 429-
435. doi:0099-1112/00/6504-429$3.00/0
Rujoiu-Mare, M. R., & Mihai, B. A. (2016). Mapping Land Cover Using Remote Sensing Data and
GIS Techniques: A Case Study of Prahova Subcarpathians. Environment at a Crossroads:
SMART approaches for a sustainable future. 32, pp. 244-255. Elsevier. Retrieved 2017
Sentinel, ESA. (2017). https://sentinel.esa.int/web/sentinel/home. Retrieved from
https://sentinel.esa.int/web/sentinel/user-guides/sentinel-2-msi/processing-
levels/level-2
Styrelsen for Dataforsyning og Effektivisering. (2017). (S. f. Effektivisering, Producer) Retrieved
March 2017, from kortforsyningen.dk: https://download.kortforsyningen.dk/
Tarabalka, Y., Benediktsson, J. A., & Chanussot, J. (2009). Spectral–Spatial Classification of
Hyperspectral Imagery Based on Partitional Clustering Techniques. IEEE Transactions on
Geoscience and Remote Sensing, 47(8), 2973-2987. doi:10.1109/TGRS.2009.2016214
Thomlinson, J. R., Bolstad, P. V., & Cohen, W. B. (1999, October). Coordinating Methodologies for
Scaling Landcover Classifications from Site-Specific to Global: Steps toward Validating
Global Map Products. Remote Sensing of Environment, 70(1), 16-28.
doi:https://doi.org/10.1016/S0034-4257(99)00055-3
United States Geological Survey. (2017, March 28). (United States Geological Survey) Retrieved
March 2017, from EarthExplorer - USGS: https://earthexplorer.usgs.gov/distribution
United States Geological Survey Metadata. (2017). Metadata. Retrieved from EarthExplorer:
https://earthexplorer.usgs.gov/metadata/10880/1174275/
Verbyla, D. (2013, November). Estimating Classification Accuracy Using ArcGIS. Retrieved May
2017, from https://www.youtube.com/watch?v=9dGjuEQie7Y&t=2s
Wang, M., Wan, Y., Ye, Z., & Lai, X. (2017). Remote sensing image classification based on the
optimal support vector machine and modified binary coded ant colony optimization
48
algorithm. Information Sciences, 402, 50-68.
doi:http://dx.doi.org/10.1016/j.ins.2017.03.027
Wehmann, A. (2013). SVM Tools for ArcGIS 10.1+. Columbus, OH. Retrieved from
http://awehmann.nfshost.com/svm.shtml
Recommended