Engineering and Technology Journal Vol. 37, Part A. No.

04, 2019
DOI: http://dx.doi.org/10.30684/etj.37.4A.5
Oday Z. Jasim Mapping LCLU Using Python Scripting
Civil Engineering Department,
University of Technology Baghdad,
Iraq. [email protected] ABSTRACT:- Land cover land use changes constantly with the time at
local, regional, and global scales, therefore, remote sensing provides
Khalid I. Hasoon wide, and broad information for quantifying the location, extent, and
Civil Engineering Department, variability of change; the reason and processes of change; and the
University of Technology Baghdad, responses to and consequences of change. And considering to the
Iraq. [email protected] importance of mapping of (LCLU). For that reason this study will focus
on the problems arising from the traditional classification (LCLU) that
Noor E. Sadiqe based on spatial resolution only which leads to prediction a thematic
map with noisy classes, and using a new method that depend on spectral
Civil Engineering Department, and spatial resolution to produce an acceptable classification and
University of Technology Baghdad, producing a thematic map with an acceptable database by using
Iraq. artificial neural network (ANN) and python in additional to other
[email protected] program. In this study the methods of classification were studied through
using two images for the same study area , rapid eye image which has
three spectral bands with high spatial resolution(5m) and Landsat 8
image (high spectral resolution with eight bands), also several programs
Received on: 08/01 /2019 like ENVI version 5.1, Arc GIS version 10.3, Python 3, and GPS. The
Accepted on: 27/02/2019
Published online: 25/04/2019 result for this research was sensuousness as geometrics accuracy
accepted in map production.

Keywords:- land use, land cover, artificial neural network, support

vector machine
How to cite this article: O.Z. Jasim, K.I. Hasoon and N.E. Sadiqe “Mapping LCLU Using Python Scripting,”
Engineering and Technology Journal, Vol. 37, Part A, No. 04, pp. 140-147, 2019.

1. Introduction spectral and spatial information to produce a

Land cover refers to the physical and biological thematic map [4]. This method also allows the
cover over the surface of the land, including water, integration of multi-sensor data, such as rapid eye
vegetation, bare soil, and/or artificial structures. and Landsat. Considering these advantages, this
Land use is characterized by the arrangements, study develops a pixel-based classification
activities, and inputs relating to people in a certain approach to produce a thematic map for the study
land cover type to produce, change, or maintain it. area using rapid eye and Landsat satellite image.
Thus, both Land Cover and Land Use (LCLU) Land cover refers to the physical and biological
maps are beneficial for many geospatial cover over the surface of the land, including water,
applications such as change detection, natural vegetation, bare soil, and/or artificial structures.
disasters, urban planning, and site selection [1]. Land use is characterized by the arrangements,
LCLU maps are often created by using remote activities, and inputs relating to people in a certain
sensing data with classification techniques. There land cover type to produce, change, or maintain it.
are wide ranges of data that can be utilized for Thus, both Land Cover and Land Use (LCLU)
classification maps, these data can be grouped into maps are beneficial for many geospatial
airborne-based and satellite-based. The accurate applications such as change detection, natural
classification methods are based on a sensor, disasters, urban planning, and site selection [1].
spatial and spectral resolutions, optical or active LCLU maps are often created by using remote
[2]. The rapid eye image is one of the optical sensing data with classification techniques. There
satellites and it has three spectral bands with high are wide ranges of data that can be utilized for
spatial resolution (5 m) [3]. Therefore, the classification maps, these data can be grouped into
captured data of Rapid Eye image has fine details airborne-based and satellite-based. The accurate
of information about a study area. However, these classification methods are based on a sensor,
data requests to robust image processing and spatial and spectral resolutions, optical or active
classification technique for extracting the features. [2]. The rapid eye image is one of the optical
The classification techniques grouped into two satellites and it has three spectral bands with high
methods such as pixel-based and object-based. The spatial resolution (5 m) [3]. Therefore, the
pixel-based method is a common classification. captured data of Rapid Eye image has fine details
Several researchers were tried to combine the of information about a study area. However, these

Copyright © 2019 by UOT, IRAQ 140

Engineering and Technology Journal Vol. 37, Part A. No. 04, 2019

data requests to robust image processing and However, still, the desired accuracies could not be
classification technique for extracting the features. achieved due to the limitations of the individual
The classification techniques grouped into two pixels. This study proposes a spatial-spectral
methods such as pixel-based and object-based. The feature extraction method based on ANN for
pixel-based method is a common classification. image classification. It is expected that the
Several researchers were tried to combine the proposed method to outperform other traditional
spectral and spatial information to produce a technique while it is computing performance
thematic map [4]. This method also allows the remains comparable to the traditional methods.
integration of multi-sensor data, such as rapid eye
and Landsat. Considering these advantages, this 3. Methodology
study develops a pixel-based classification In this research, datasets of Landsat and Rapid Eye
approach to produce a thematic map for the study images are going to be combined with proper
area using rapid eye and Landsat satellite image. spectral merging techniques such as neural
networks. The spectral information of the image
2. Problem Statement will be used using the reflectance values, while the
Integration of different multispectral data can lead spatial information will be extracted by a series of
to an improved LCLU if proper classification convolutional filters. Then, the combined features
methods are applied [5]. However, this integration will be used to predict the label of each individual
also creates a challenge in regards to computing pixel in the area and produce an LCLU map. The
performance. Thus, many researchers have final map then will be produced by post processing
attempted to improve the traditional classification the created LCLU map in the previous step.
methods (e.g., maximum likelihood) [6]. Attempts Finally, a proper evaluation method will be applied
were made to address the challenges above by to measure the improvements made by our method
selecting suitable algorithms to merge data from over other methods. As illustrated in the flowchart
different sensors and then run a classification 1.
algorithm on the compressed data. This has
resulted in better classification accuracies and
better computing performances.

Landsat 8 Rapid Eye

Tools:-
 ENVI
Image Pre-processing  GIS
• Radiometric Calibration Geometric  Python
correction Atmospheric correction

Support vector machine Model

(SVM) using ENVI

Suitable algorithm Artificial

Neural Network Model (ANN))
using python
Classification
based on Rapid
Classification based on
Landsat 8 and Rapid Eye Assessment
together the
Accuracy

Base Map

Flowchart 1: The sequence of operation

141
Engineering and Technology Journal Vol. 37, Part A. No. 04, 2019

4. Research Objectives automatic memory management. It supports

The primary goal of this research is to production multiple programming paradigms, including
Land cover Land use for the study area based on object-oriented, imperative, functional and
ANN for image classification. The specific procedural, and has a large and comprehensive
objectives are: standard library. Python interpreters are available
1. To develop a model based on ANN for image for many operating systems. CPython, the
classification. reference implementation of Python, is open
2. To compare the accuracy and computing source software and has a community-based
performance of the proposed method with the development model, as do nearly all of Python's
traditional support vector machine and other other implementations. Python and CPython are
methods. managed by the non-profit Python Software
Foundation.
5. Experimental Work
IV. Artificial Neural Network Model(ANN)
I. Study Area
Neural network is a machine that is designed to
This study is limited to the evaluation of LCLU in model the way in which the brain performs a
Karbala for future design and planning. The particular task or function of interest; the network
proposed model is easy-to-use engineering is usually implemented by using electronic
methods for image classification. In addition, the components or is simulated in software on a digital
LCLU maps generated by ENVI, Python, and GIS computer [7], it has several tasks such as
can be available by which decision-makers can classification, regression, clustering,
check the situations before and after the planning Dimensionality reduction and others. These tasks
for the study area. depend on the desired output of a machine-learned
The research covers the area (25 km) in Karbala.
system [8]. In general ANN Consists of (input,
In this study used two images: i) Rapid Eye image hidden, and output layers) that receive process and
has three bands with 5 m panchromatic resolution present the results [9] as shown in Figure 2.
and ii) Landsat 8 image has eight bands with 15 m
panchromatic resolution. Used these data to
acquire valuable information from the proposed
model for producing LULC map. Artificial Neural
Network (ANN) was utilized the proposed
predicted model for Thematic map. As shown in
Figure 1.

II. Significance of the Study

This study attempts to provide better insights into
a pixel-based classification for multi-sensor data.
It examines the advantages of integrating Landsat
and Rapid Eye images to provid LCLU maps. It
also analyzes some advanced classification
methods such as ANN to analyze such datasets
with less computation time. Overall, this study will Figure 1: shows the area of study
improve the understanding of general users, urban
planners, and decision-makers about generating
LCLU maps from integrated remote sensing data.

III. Python
Python is an interpreted high-level programming
language for general-purpose programming,
created by Guido van Rossum and first released in
1991; Python has a design philosophy that
emphasizes code readability, notably using
significant whitespace. It provides constructs that
enable clear programming on both small and large
scales. In July 2018, Van Rossum stepped down as
the leader in the language community after 30
years. Python features a dynamic type system and Figure 2: A simple structure of the neural network

141
Engineering and Technology Journal Vol. 37, Part A. No. 04, 2019

ANN model has two equations, which are 6. Process and Result Obtained from the Work
considered very important to improve the accuracy Approach
of the ANN model. First equation for minimizing Satellite Data that used in this study as shown in
the error function as illustrated in (Equation 1), the Table 1, the Landsat 8 image has eight spectral
second equation for corrections to weight bands with 30 m spatial resolution which covered
parameters such as computed and added to the visible to near-infrared and one panchromatic band
previous values (Equation 2). with a spatial resolution of 15 m [12], and Rapid
Eye has three spectral bands with 5 m spatial
𝐿 resolution The spectral regions of bands are Blue
1 2
𝐸 = ∑(𝑑𝑗 − 𝑜𝑗𝑀 ) (1) (440-510 nm), Green (520-590 nm), Red (630-
2 690 nm) [13]. Figure 3 illustrated the Landsat and
𝑖=1
rapid eye images. These two images are geo-
Where, referenced to follow the Universal Transverse
𝑑𝑗 Represents the desired output. Mercator projection (UTM Zone 38N) and the
𝑜𝑗𝑀 Represents the current response of the node in world geodetic system 1984 datum (WGS 1984).
the output layer. In addition, the images need to correct the
L is the number of nodes in the output layer. radiometric correction and before the performance
of the classification. Image Pre-processing
𝜕𝐸 Radiometric Calibration, Geometric correction,
∆𝑤𝑖,𝑗 = −𝜇 and Atmospheric correction using ENVI. Pan
{ 𝜕𝑤𝑖,𝑗 (2)
Sharpening algorithm was used to integrate the
∆𝑤𝑖,𝑗 (𝑡 + 1) = ∆𝑤𝑖,𝑗 + 𝛼∆𝑤𝑖,𝑗 (𝑡) multispectral image (Landsat 8) with a high spatial
resolution image (Rapid Eye) to acquire the
Where, improving visualization of urban and other
𝑤𝑖,𝑗 is weight parameter between node 𝑖 features. As illustrated in Figure 4 the study area
𝑗, ∆ represents a learning rate after these processes. Selecting the location of the
𝛼 Represents a momentum factor t, the parameter samples. In this study was used geographic
𝛼can be called smoothing or stabilizing factor as it information system (GIS) to select these samples
smoothest the rapid changes between the weights through create random point algorithm? This
[10]. algorithm worked through identifies the boundary
of the study area and select the sample percentage
V. Support Vector Machines model for the whole study area. After applying to create
Support Vector Machine (SVM) was first heard in random points of the area, drawing the features and
1992, Support vector machines (SVMs) are a set identify the type of features such as grass,
of related supervised learning methods used for building, water and others. As shown in Figure 5.
classification and regression. Then using ARC GIS to convert Pan Sharpening
SVM is a statistical classification method that have image from raster to point, convert the point to the
yi |x⃑⃑⃑i ∈ Rn , yi ∈
m labeled training samples, {(x⃑⃑⃑i , ⃑⃑⃑ table. Classifying these tables in ANN using
{−1,1}, i = 1 … m, SVM can generate a separation python program.
hyper surface with maximum generalization
ability. Mathematically, the decision function can Table 1: Details of the datasets
be formulated as follows: Dataset Date of Spatial Spectral
Acquisition Resolution Resolution
m
Landsat 28/04/2013 15 m 7 bands
⃑⃑⃑⃑⃑ = ∑ αi yi K(x⃑⃑⃑i , x⃑) + b
d(x) (3) 8
i=1 Rapid Eye 28/04/2012
5m 3 bands
Where αi and b are the parameters determined by
the SVM learning algorithm, and K(x⃑⃑⃑i x⃑) is the
kernel function, which implicitly maps the samples
to a high dimension space. The samples ⃑⃑⃑ xi with
nonzero parameters αi is called support vectors.

141
Engineering and Technology Journal Vol. 37, Part A. No. 04, 2019

Figure 3: shows Rapid Eye image (left) and Landsat 8 image (right).

I .Results of ANN
The Artificial Neural Network (ANN) was used
with the best hyper parameters. This model was
used to classify the combined image (the Landsat
with Rapid Eye images) into five land cover
classes such as urban, road, bare land, water, and
agriculture. The overall accuracy and the kappa
accuracy of the ANN model were found to be
0.983 and 0.976, respectively for on the training
dataset. On the other hand, the accuracy of the
testing dataset was 0.951 and 0.942 for overall
accuracy and the kappa accuracy respectively.

II Result of SVM
In this study compared our work with another
method like support vector machines (SVM). The
Figures 4: shows Sharpening image the Rapid Eye with SVM method with the best hyper parameters for
Landsat images these methods, the overall performance SVM
model was about 0.91. Figure 6 shows the Rapid
Eye, the sharpening images (Landsat with Rapid
Eye images) and the classification maps of SVM
and ANN classifiers. Regarding the classification
map of SVM, the water, the roads, and the bare
land classes are very noisy and are not clear. While
the classification map of ANN classifier with
sharpening image seems clearer and more logical.
The results indicated firstly the ANN model is
better than the SVM model as well as the
multispectral image better than Rapid Eye image.
The multispectral image is better because it has
eight bands with high spatial resolution 5 meters.
Nevertheless, the Rapid Eye image has three bands
with high spatial resolution 5 meters. It is obvious
the spectral band is very important for analysis to
produce the Land use and land cover mapping
besides the high spatial resolution.
Figure 5: shows the samples for the study area.

144
Engineering and Technology Journal Vol. 37, Part A. No. 04, 2019

𝑁 ∑𝑟𝑖=1 𝑥𝑖𝑖 − ∑𝑟𝑖=1(𝑥𝑖 + 𝑥+𝑖 )

𝐾 = (4)
𝑁 2 − ∑𝑟𝑖=1(𝑥𝑖 + 𝑥+𝑖 )

Where 𝑟 is the number of rows in the matrix, 𝑥𝑖𝑖 is

the number of observations in row 𝑖 and column 𝑖,
𝑥𝑖 and 𝑥+𝑖 are the marginal totals for row 𝑖 and
column 𝑖 respectively and 𝑁 is the total number of
pixels.
In this research, the Accuracy Assessment for
ANN& SVM are 0.97,and 0.91respectively. Table
2 shows the summary of the accuracy assessment
for SVM and ANN methods.
Finally the accuracy assessments and sensitivity
analysis of ANN Model
1. Network Analysis:-In this study was determined
the hidden layers that affected the performance of
network analysis depth. And also was checked
Figure 6: shows the classification maps produced several network depth ranging from 10 to 100
from support vector Machines and Artificial
hidden layers. The evaluation of the ANN model
Neural Network Model method.
was based on kappa accuracy and overall accuracy
III. Accuracy Assessment that was measured on training and testing datasets.
The Table 3 was presented the summary of the
In LCLU classification, accuracy assessment of results
the classification results is often required. The
classification maps as created from remote sensing Table 2: The accuracy of ANN and SVM classifiers
images may contain many errors due to many Overall accuracy kappa accuracy
reasons including geometric error, complete Classifi
er Traini Testin Traini Testin
atmospheric corrections, or due to ng g ng g
indistinguishable classes. This is why accuracy
assessment procedure is very important. Based on ANN
0.974 0.954 0.959 0.936
Model
the confusion matrix analysis, the overall accuracy
and a Kappa analysis were used to get the
classification accuracy assessment. The overall SVM
0.917 0.884 0.873 0.851
Model
accuracy was computed by dividing the total
correct (some of the major diagonal) by the total
number of pixels in the error matrix. However, the
Kappa, which is a discrete multivariate technique .
used in accuracy assessments, which was
computed as:

Table 3: Results of network depth analysis.

Experiment Hidden Overall accuracy kappa accuracy
No. Layers Training Testing Training Testing
1 10 0.661 0.623 0.522 0.485
2 15 0.681 0.623 0.632 0.545
3 25 0.693 0.635 0.644 0.557
4 40 0.711 0.653 0.662 0.575
5 60 0.753 0.694 0.701 0.635
6 75 0.798 0.771 0.783 0.758
7 85 0.852 0.837 0.809 0.784
8 95 0.896 0.876 0.834 0.814
9 100 0.974 0.954 0.959 0.936
10 105 0.861 0.846 0.814 0.807
11 110 0.807 0.793 0.787 0.769
12 120 0.694 0.669 0.658 0.608

141
Engineering and Technology Journal Vol. 37, Part A. No. 04, 2019
DOI: http://dx.doi.org/10.30684/etj.37.4A.5
2. Optimization Algorithm Analysis:- The SGD and ADAM. Table 4 summarized the best
optimizations of ANNs is very important because accuracy, and the suitable algorithm for optimizing
lead to improve the performance of the network ANN classifier to classify the Sharpening image.
model. There are many algorithms and several
ways help to develop this network such as LBFGS,

Table 4: Results of the optimization algorithm analysis.

Experiment Overall accuracy kappa accuracy
Solver
No. Training Testing Training Testing
1 SGD 0.673 0.668 0.663 0.659
2 ADAM 0.872 0.856 0.839 0.828
3 LBFGS 0.974 0.954 0.959 0.936

3. Activation Function Analysis for ANN The Tables 5 and 6 show the summary of the
modeling:-Finally, the activation function was activation function analysis for hidden layers
considered important for both the hidden and and output layers respectively
output layers because of this function influence
on the overall performance of ANN modeling.

Table 5: Results of the activation function analysis for hidden layers.

Experiment Overall accuracy kappa accuracy

Activation
No. Training Testing Training Testing
1 IDENTITY 0.841 0.829 0.806 0.792
2 LOGISTIC 0.761 0.753 0.742 0.735
3 TANH 0.974 0.954 0.959 0.936
4 RELU 0.898 0.884 0.867 0.853

Table 6: Results of the activation function analysis for output layers.

Experiment Overall accuracy kappa accuracy
Activation
No. Training Testing Training Testing
1 IDENTITY 0.857 0.836 0.817 0.804
2 LOGISTIC 0.789 0.761 0.754 0.739
TANH
3 0.886 0.872 0.854 0.837

4 RELU 0.974 0.954 0.959 0.936

7. Conclusions 3. The accuracy of the classification map was

1. When compared the results of ANN model with depended on the testing and training samples. The
results of SVM, the former model achieved high experiment number 9 with 100 hidden layers was
classification accuracies and the maps are more achieved the best accuracy. On the training and
logical, it is mean the performance of ANN testing, dataset the overall accuracy and kappa
classifier outperforms the SVM method with best accuracy were 0.974, 0.954 and 0.959, 0.936
hyper parameters for these methods. The overall respectively. The results were indicated there are
performance of the ANN model was about 0.97 no specific patterns or number of hidden layer to
while the SVM model was about 0.91, so this study improve the ANN model.
suggests that the ANN model with these hyper 4. For Optimization, Algorithm Analysis found
parameters could significantly outperform the that the experiment number 3 with solver LBFGS
SVM model for Karbala area. As well as, was the suitable algorithm for land cover mapping,
regarding the classification map of the ANN model as well as the same classifier achieved the QA and
was more logical and better than the classification K were 0.974, 0.954 and 0.959, 0.936 respectively.
map of SVM. 5. In addition, for Activation Function Analysis
2. The spectral band is very important for analysis for ANN modeling, this model analyzed several
to produce the Land use and land cover mapping activation functions such as Logistic, RELU,
besides the high spatial resolution. identity and TANH. Therefore, the analysis of the

Copyright © 2019 by UOT, IRAQ 140

Engineering and Technology Journal Vol. 37, Part A. No. 04, 2019

ANN shows the best accuracy was achieved with https://www.satimagingcorp.com/satellite-

TANH and, RELU for hidden layers and, output sensors/other-satellite-sensors/rapideye/
layers respectively, while the worst result of the
ANN model was logistic activation function used.

References
[1] S. Liang, X. Li and J. Wang, “Advanced Remote
Sensing,” Academic Press is an imprint of Elsevier, 1st
ed. USA, p.704, 2012.
[2] Q. Weng, “Remote Sensing and GIS Integration
Theories, Methods, and Applications,” The McGraw-
Hill Companies, pp12-21, 2010.
[3] G. Metternicht, L. Hurni, and R. Gogu, “Remote
sensing of landslides: An analysis of the potential
contribution to geo-spatial systems for hazard
assessment in mountainous environments,” Remote
Sensing of Environment 98, pp. 284–303, 2005.
[4] C. Cleve, M. Kelly, F. R. Kearns and M. Moritz,
“Classification of the wildland–urban interface: A
comparison of pixel- and object-based classifications
using high-resolution aerial photography,” Computers,
Environment and Urban Systems 32, pp. 317–326, 2008
[5] D. Lu, and Q. Weng, “Urban Classification Using
Full Spectral Information of Landsat ETM Imagery in
Marion County, Indiana,” Photogrammetric
Engineering & Remote Sensing Vol. 71, No. 11, pp.
1275–1284, 2005.
[6] P. Bolstad and T.M. Lillesand, “Rapid maximum
likelihood classification,” Photogrammetric
engineering & remote sensing, Vol. 57, No. 1, pp. 67-
74, 1991.
[7] S. Haykin, “Neural Networks and Learning
Machines,” Pearson Prentice Hall 3rd ed. USA, p.2,
2009.
[8] M. Mokhtarzade and M. V. Zoej, “Road detection
from high-resolution satellite images using artificial
neural networks,” International journal of applied earth
observation and geo-information, Vol. 9, No. 1, pp.32-
40, 2007.
[9] J. Gao, “Digital Analysis of Remotely Sensed
Imagery,” McGraw-Hill Companies, 1st ed. USA,
pp.306-307, 2009.
[10] J. Gao, “Digital Analysis of Remotely Sensed
Imagery,” McGraw-Hill Companies, 1st ed. USA, p323,
2009.
[11] A.A. Matkan, M. Hajeb and S. Sadeghian, “Road
Extraction from Lidar Data Using Support Vector
Machine Classification,” Photogrammetric Engineering
& Remote Sensing Vol. 80, No. 5, pp. 409–422, 2014.
[12] K. Zanter, “LANDSAT 8 (L8) DATA USERS
HANDBOOK,” Department of the Interior U.S.
Geological Survey, 3ed, 2018 Available on:
https://landsat.usgs.gov/landsat-8-l8-data-users-
handbook-section-
2?fbclid=IwAR0quSRlxR_StY5hhzDX-
i8OWKP1e5bqu5lpZOSTbAbVsZr7kmbvR4tap1k
[13] B. Bridge, “Rapid Eye Sensor,” Satellite Imaging
Corporation 2017 Available on:

141