Journal of Geosciences and Geomatics, 2018, Vol. 6, No.

3, 103-123
Available online at http://pubs.sciepub.com/jgg/6/3/2
©Science and Education Publishing
DOI:10.12691/jgg-6-3-2

UAV-based Approach to Extract Topographic and

As-built Information by Utilising the OBIA Technique
Hairie Ilkham Sibaruddin1,2, Helmi Zulhaidi Mohd Shafri1,2,*, Biswajeet Pradhan3, Nuzul Azam Haron1
1
Department of Civil Engineering, Faculty of Engineering, Universiti Putra Malaysia (UPM), 43400 Serdang, Selangor, Malaysia
2
Geospatial Information Science Research Centre (GISRC), Faculty of Engineering, Universiti Putra Malaysia (UPM),
43400 Serdang, Selangor, Malaysia
3
School of Information, Systems & Modelling, Faculty of Engineering and IT, University of Technology Sydney (UTS),
Sydney, Australia
*Corresponding author: helmi@upm.edu.my; hzms04@gmail.com

Received August 08, 2018; Revised September 19, 2018; Accepted October 07, 2018
Abstract In this study, the capability of Unmanned Aerial Vehicle (UAV) optical data to provide reliable
topographic and as-built information was tested using the eBee Sensefly UAV system. The Object-based Image
Analysis (OBIA) technique was used to extract important geospatial information for mapping. The robust Taguchi
method was adopted to optimise the segmentation process. Feature space optimisation method was used to obtain the
best features for image classification utilising different supervised OBIA classifiers, such as K-nearest neighbour
(KNN), normal Bayes (NB), decision tree (DT), random forest (RF) and support vector machine (SVM). Results
showed that SVM obtained the highest percentage of overall accuracy, followed by RF, NB, DT and KNN at
97.20%, 95.80%, 93.14%, 86.01% and 77.62%, respectively. The McNemar test was implemented to analyse the
significance of the classifier results. The as-built information showed that dimensional accuracy was less than 1
metre compared with ground survey measurement. We conclude that the combination of UAV and OBIA provides a
rapid and efficient approach for map updating. This technique could replace the current procedure that utilises
piloted aircraft and satellite images for data acquisition and reduce the time for digitising each feature that represents
land cover for urban mapping.
Keywords: UAV, land cover, topography map, as-built, OBIA, segmentation
Cite This Article: Hairie Ilkham Sibaruddin, Helmi Zulhaidi Mohd Shafri, Biswajeet Pradhan, and Nuzul
Azam Haron, “UAV-based Approach to Extract Topographic and As-Built Information by Utilising the OBIA
Technique.” Journal of Geosciences and Geomatics, vol. 6, no. 3 (2018): 103-123. doi: 10.12691/jgg-6-3-2.

given the constraint and requirement to map small areas

and using large-format aerial or metric cameras to acquire
1. Introduction data is uneconomical and unsuitable [1].
Unmanned Aerial Vehicle (UAV) system operates
A topographic map provides important information that a powered aerial vehicle without a human operator.
represents data on land use and land cover for certain UAVs are prominent due to their provision of data with
areas. It is a 2D representation of the Earth’s 3D landscape. high spatial resolution [2], lightweight sensors and
Topographical data also provide an accurate measured platforms, flexibility of flight planning and deployment
plan of a site that encompasses the entire range of various and elimination of long dependency [3]. UAVs could also
feature information with detailed illustration of man-made obtain timely imagery of areas that are dangerous or
and natural features on the ground, such as road, railways, difficult to access by traditional means. This imagery can
rivers, lakes and buildings. Typically, a topographic map usually be acquired at a minimum cost or at a cost that is
is used as a skeleton for design work before a construction cheaper than that involved in other collection methods
project begins to address the requirements of land survey, [2,3,4]. Current users prefer technologies with low cost
urban planning, as-built planning, hazard assessment and but numerous benefits. UAVs are an example of such
disaster risk management. technologies because they provide highly applicable,
In the time frame of a mapping survey, a dataset of the immediate and near-real-time data at a resolution that is
topographic map is gathered from platforms, such as comparable to that of terrestrial means. UAVs, such as
space-borne satellites and manned aircraft. Most of the eBee Sensefly are an excellent technology that can
data are acquired by equipment that are too expensive to provide high capability data for mapping purposes [5].
build and maintain for small-area map updating. In The features from aerial photo orthomosaic are normally
addition, the data are not always within the public domain. detected and digitised manually from visual interpretation
The process of acquiring aerial mapping is expensive for mapping purposes. However, these methods consume
 Journal of Geosciences and Geomatics 104

much time, are tedious and expensive [6]. Automation will such as satellite image and aerial laser scanner, for
provide substantial benefits [4]. The level of automation building extraction [21,22,23,24].
could range from semi-automatic incorporated with human Therefore, the first objective of this study is to assess
interaction to completely automated [7]. The potential the capability of UAV to provide reliable topographic and
of acquiring accurate and low-cost UAV data relies on as-built data information by utilising the OBIA technique.
automatic object reconstruction and boundary extraction Specifically, the aim is to determine the most optimal
activities [8]. OBIA parameters through segmentation and classification
Pixel-based images analysis are often used to extract to deliver the required information from UAV data. The
low-level features. However, an image is classified segmentation process is important for object classification
according to spectral information, and the pixels in the intended for object-oriented image analysis. This study
overlapping region are misclassified; the salt-and-pepper investigated the effect of parameter tuning with different
problem thus emerges in the classification result [9] sample numbers on the overall accuracy of the results
and causes confusion among classes [10]. Meanwhile, to determine the optimal parameter. Machine learning
Object-based Image Analysis (OBIA) is used to extract classifiers were used. The second purpose is to extract
high-level features, which constitute shapes in images topographic information, such as land cover features,
that are detected regardless of illumination, translation, from UAV data. Lastly, the study aims to extract
orientation and scale [7]. as-built information, such as infrastructure geometry and
Object-based classification for high-spatial-resolution dimensions. The geometry from OBIA data was compared
UAV data encounters several challenges despite its highest from ground truth survey data using a high-accuracy total
accuracy among all types of sensors [11]. The scale station equipment.
parameters used for segmentation are much larger than those
used for aerial and satellite imagery. The extreme detail on
the imagery parses it into many different objects with 2. Data and Methods
varying spectral, morphological and proximity characteristics.
Most segmentation processes use trial and error, which 2.1. Study Area
is subjective, laborious and time consuming [10]. Hence, a
solution to optimise the segmentation process for The study area is situated at the National Land and
classification is required. The Taguchi method, which was Survey Institute (INSTUN) in Behrang, Ulu, Tanjung
designed by Dr. Genichi Taguchi, has a simple statistical Malim, Perak, Malaysia. The total area of this campus is
tool design [12]. This method involves a tabulated design approximately 200 acres. The area of research interest is
(arrays) system that permits a maximum number of main limited between latitudes 3° 45’ 58.3’ N to 3° 46’ 2.16’ N
effects to estimate in an unbiased manner with the lowest and longitudes 101° 30’ 34.94’ E to 101° 31’ 26.01’ E,
number of experimental runs [13]. Several studies had with a total area of 0.3628 km2. The study area is
applied this method to optimise the segmentation process surrounded by man-made infrastructures, such as buildings,
in OBIA [10,14,15,16]. roads, drainages, sport courts, concrete benches, pavements
Most of the studies have utilised UAV data to produce and parking lots. Natural features, which are dominant,
topographic (DTM, DSM, orthophoto) [17] and land cover include bare soil, dead grass, grass lands, sand, crops,
maps [11,18,19,20]. However, the data are not fully shrubs and trees. Other features include water bodies, such
utilised for as-built plan information. A few studies have as swimming pool, lakes, septic tanks and shadows from
highlighted infrastructure information using other sensors, tall buildings and trees.

Figure 1. Study area in INSTUN, Perak, Malaysia

105 Journal of Geosciences and Geomatics

2.2. Methodology involved image segmentation, selection of training and

testing samples, image classification, feature selection,
The methodology of this research was divided into three tuning parameter setting for each classifier and accuracy
phases as shown in Figure 2. The initial phase relied on assessment, was performed. All related data, such as digital
the acquisition of data from the eBee Sensefly UAV. The surface model (DSM), digital terrain model (DTM), contour
orthorectified images were generated using photogrammetric line, image classification output for generating topography
technique. Then, an object-based image analysis, which map and as-built plan and information, were combined.

Figure 2. Research methodology flowchart

 Journal of Geosciences and Geomatics 106

2.2.1. Phase 1 - Data Acquisition: Pre-processing of

UAV Image
The imagery data were obtained on October 18, 2016
using eBee Sensefly UAV. The camera sensor, which is
attached on this model, is Canon 16 MP IXUS with a
visible colour band (red, green, and blue). During data
acquisition, the side lap was 60%, and 80% of the
front overlap had been set. The altitude of the UAV was
set to 190 m above the ground level. The data were tied
with six control points (benchmark and EDM calibration
pillars).
The quality check on georeferencing showed that the
mean RMS error was 0.025 m. The entire data were set
geometrically with the coordinate system WGS84 datum
and 47N zone in UTM projection. Raw images were
mosaicked to generate an orthorectified image that Figure 3c. DSM
covered the entire study area by using photogrammetry
software Pix4D. The average ground sampling distance
(GSD) through this orthorectified image was 5 cm. Seven
classes of features were organised and investigated as
follows: (1) soil/sand, (2) urban tree, (3) building/roof,
(4) impervious surface (other infrastructures), (5) grassland,
(6) water body and (7) shadows.
An orthomosaic image with the DSM, DTM and
contour line was generated (Figures 3(a–d)) using Pix4D.
The image was subjected to automatic radiometric and
geometric correction.

Figure 3d. Contour Line (1m)

2.2.2. Phase 2 - Segmentation and Classification

The initial and most important process during the
implementation of the OBIA technique is the segmentation to
divide an image into meaningful sections that are correlated
to objects in the real world, as shown in the image [25].
Multi-resolution segmentation is a type of bottom-up,
region-based segmentation algorithm [26] and is applied
using the software e-Cognition version 9.0. Image
classification accuracy is promptly controlled with the
Figure 3a. Orthomosaic
quality of the segmentation, which is in turn controlled by
defined parameters [27]. The three parameters in the
multi-resolution segmentation process are as follows:
scale, shape and compactness.
The most effective parameter that affects the average
image object size is the scale factor [25,28,29]. This factor
is controlled by the spatial resolution of the image and
features [9,12]. Shape and compactness factors are
associated with colour density and smoothness. Then, the
amount of spectral information that should be aggregated
to build the segments is identified [14]. Initially, to obtain
the best range of the scale parameter for image
segmentation, a trial and error process was conducted in
this work. The selected scale on segmentation was further
optimised by applying the Taguchi method.
The segmentation process began with defining the
possible range of multi-resolution segmentation parameters
Figure 3b. DTM to identify the pertinent value of the scale and homogeneity
107 Journal of Geosciences and Geomatics

parameter. The analysis was performed with different Intra-segment variance,

scale factors at 5, 25, 50, 75, 100, 125, 150, 175, 200 and
n

∑a v
225. Visual interpretation was performed to identify the
reliability of the segmented image. The main criterion was i i

that the image object was under-segmented, and the v= i =1

n
(1)
over-segmented image object is eliminated. Previous
studies [17,25,30] have consistently emphasised the scale
∑a
i =1
i

parameter while the other factors remained constant. The

shape parameter was set to 0.1, and compactness was set where ɑi and vi refer to area and variance respective to
to 0.5 in all 10 preliminary tests to generate meaningful region i, respectively. Intra-segment variance v is a weighted
segmented objects. Both parameters used these default average, where the weights are the areas of each region.
values for image segmentation. The second element assessed is intersegment heterogeneity.
The five best results of preliminary segmentation were The function employs Moran’s I autocorrelation index
identified via visual interpretation. The advantages of the [36]. It quantifies the degree of spatial affiliation as
Taguchi method include minimising the number of reflected in the data as a unit [34].
experiments [15] by adopting a fractional functional design Moran’s I index is expressed as follows:
and maintaining a consistent and simple experimental n n
design. This system approach can significantly reduce the n∑∑ wij ( yi − y )( y j − y )
=i 1 =j 1
total testing time and experimental cost [31]. The Taguchi I= (2)
 n 2
 ∑ ( yi − y )  ( ∑ i ≠ j ∑ wij )
orthogonal array disperses the parameter equally, and the
column depicts independent orthogonal variables to
guarantee an impartial comparison of all variables in each  i =1 
level and to examine each parameter separately [32]. The where n represents the total number of regions and wij is
experiment uses an orthogonal array within each pair of the spatial weight between objects i and j. yi is the mean
columns that corresponds as independent variables. Level grey value of region Ri, and ȳ is the mean grey value of
combination exists in an equal number of times [13,14,33]. the image (1 for adjacent regions and 0 otherwise). The
The orthogonal array design by Taguchi is limited to test was executed, and the corresponding plateau objective
only the combinations of 25 experiments with 3 varying of function (POF) was computed for each test based on the
main parameters rather than considering 125 (5x5x5) combination of parameter level in the orthogonal arrays.
experimental probabilities. The Taguchi orthogonal array The test with the highest result of POF revealed the best
was applied here using Minitab v.17 software. Prior to performance and was marked as the strength of the quality
undertaking the statistical Taguchi optimisation, five of optimisation [34]. The objective function (F) was
levels of the three parameters were defined, as illustrated combined with the within-segment variance (v) measure
in the following table. and the between-segment autocorrelation of Moran’s I
index (I) [34]. It can be expressed as follows:
Table 1. Level for segmentation parameters
LEVEL F (v=
, I) F(v) + F (I) (3)
PARAMETER
1 2 3 4 5 where F(v) and F(I) are the normalisation functions.
Scale 25 50 75 100 125
X max − X
Shape 0.1 0.3 0.5 0.7 0.9 F (x) = X max − X min (4)
compactness 0.1 0.3 0.5 0.7 0.9
Subsequently, the signal-to-noise (SN) ratio was employed
to model the optimal segmentation parameters. The three
Then, the statistical Taguchi method and the spatial types of SN ratio analysis were applied as follows:
objective function were fused in a particular process to (1) lower is better (LB), (2) nominal is the best and
optimise the segmentation parameters [33,34]. The idea of (3) higher is better (HB). The aim of this experiment was
combining statistical and spatial (objective function) to optimise the segmentation parameter for image data.
optimisation methods in a particular process is to model Hence, the HB category of the SN ratio was used
the optimal parameters that guarantee an acceptable for modelling. The SN ratio for each experiment
quality of segmentation [14]. The objective function is was calculated using mean value yi and variance si
accomplished with the fusion of spatial autocorrelation by determining the effect of each variable. It can be
and variance indices to identify relevant segmentation [35]. expressed as follows [14]:
Spatial autocorrelation implies the level of distinctiveness
2
between regions (heterogeneity). The variance indicator y
shows the uniqueness (homogeneity) of the pixels in a SN i = 10 log10 i2 (5)
single segment [34]. Hence, the condition of good-quality si
segmentation with the consequences of intra-segment
where ȳ is the mean and si refers to variance as denoted by
homogeneity and inter-segment heterogeneity is strictly
the equation.
maximised [14].
The first element computed is the intra-segment variance 1 Ni
of the regions created by a segmentation algorithm using
the equation 1 [34].
yi = ∑ yi,u
N i u =1
(6)
 Journal of Geosciences and Geomatics 108

Five machine learning classification algorithms that are

1 Ni
=si ∑ ( yi,u − yi )
N i − 1 u =1
2
(7) K-nearest neighbour (KNN), normal Bayes (NB), decision
tree (DT), random forest (RF) and support vector machine
(SVM) were used and tested thoroughly to evaluate their
In Equations (6) and (7), i is the number of experiments, performance under varying conditions and to optimise
u is the trial number and Ni is the number of trials in their applicability in terms of the OBIA technique. To
experiment i. The average SN ratio was employed to optimise the parameter for each classifier, several tests
evaluate the result of each experiment. A high SN ratio were performed by tuning the parameter for each classifier,
denotes the optimal parameter segmentation based on except for NB with one available parameter for adjustment.
Equation (8). The average of the SN value for each level The sensitivity of each classifier was examined using the
and factor was derived. Then, the result was exported as a selected training and testing samples by referring to the
table and a graph. results of the accuracy assessment and by varying their
respective parameters.
1 Ni
1 
SN i = − log10  ∑y 2
(8)
NB is a simple technique for constructing classifiers by
applying the Bayes theorem (Bayesian statistics) [37]. It is
 Ni u =1 u 
not a single algorithm for training classifiers; it is a family
Seven classification classes were identified and are of algorithms based on a common principle. The NB
illustrated in Table 2. classifier assumes that the value of a particular feature is
The training and testing samples were selected randomly independent. The data distribution function with one
based on experience from ground truth assessment data on component per class is assumed to be a Gaussian mixture.
the study area. The sample was divided into five parts to check The algorithm estimates the mean vectors and covariance
the classification accuracy and to examine the influences matrices of the selected features for each class for
of classifiers on the sample size; 70% of the sample was classification. The advantage of the NB classifier is that it
used for training and 30% for testing. Each training and does not require any parameter to tune, which could be
testing sample was selected differently to ensure that no subjective and time consuming.
repeated sample was selected for accuracy assessment. The KNN algorithm is a method for classifying objects
An assessment using different numbers of samples for based on closest training examples in the feature space.
training and testing was also conducted by tuning the KNN is a non-parameter algorithm for instance-based
parameter for each classifier while excluding NB to obtain learning or lazy learning [38]. An object is classified by
the relationship with the various parameters being set for referring to the class attributes of its KNN parameters.
optimising the classification result. The different numbers Therefore, K is the key tuning parameter in this classifier,
of samples for each class were selected randomly due to and it is largely determined the performance of the KNN
different areas for each class. The soil/sand class with the classifier [37]. In this study, the K values varied from 5 to
highest number of object classes was selected, followed 20 with 5 intervals each to identify the optimal K value for
by urban tree, building/roof, grassland, impervious surface, all training sample sets.
water and shadow, as shown in Table 3. DT learning is a method used in data mining, in which
After selecting the training sample, spectral and spatial a series of decisions are made to segment data into
features were required to classify the image. Feature space homogeneous subgroups. The aim is to create a model that
optimisation (FSO) tools, which are available in e-Cognition predicts the value of a target variable based on several
software, were used for feature selection. More than hundreds input variables. This process is repeated on each derived
of features are available for classification. A total of 41 subset in a recursive manner (recursive partitioning). The
features were selected for FSO to identify the appropriate recursion is completed when the subset at a node has the
features for further classification [18]. The object features same value as the target variable or when splitting no
were used and processed for further analysis with different longer adds value to the predictions. The purpose of
training and testing sizes. The features were divided into analyses via tree-building algorithms is to determine a set
shape, texture and spectral properties. The values of best of if–then logical (split) conditions [38]. During this study,
separation distance for samples 1, 2, 3, 4 and 5 were 3.975, we tested the value of maximum depth from 1 to 20 for all
2.545, 2.166, 2.063 and 1.882, respectively. Sample 1 five training samples. The other parameters, such as cross
with 100 training data samples had the highest separation validation folds and min sample count, were set to 10
distance compared with the other data samples. (default). Other factors remained constant.

Table 2. Types of image classification

NO. NAME OF CLASSES DESCRIPTION

1 Soil/Sand All bare sand/soil and/or very/dead grasslands
2 Urban Tree All type of crop, shrub, bush, oil palm tree species canopies
3 Building/Roof All different size building or roofs with different material cover
4 Grass Land All type of grass lands
All impervious surface areas e.g. road, park lots, pavement, tennis court, volleyball court,
5 Other Infrastructures (Impervious Surface)
bench, jetty, drainage, sump
6 Water All different types of water bodies (lake, swimming pool, septic tank)
7 Shadow All shadows from building, tree (light and dark shadow)
109 Journal of Geosciences and Geomatics

RF, also referred to as random tree, is an ensemble of classification and regression analysis. The most frequently
decision trees [37]. The random tree classifier is more of a used types of kernel functions or SVM algorithms are
framework than a specific model. It uses an input feature linear, polynomial and radial basis functions (RBF) and
vector and classifies it with every tree in the forest. sigmoid kernels [37]. In this study, the RBF kernel, which
It results in a class label of the training sample in the is the most frequently used and has been proven superior
terminal node, where it ends [38]. The value of the to other kernels, was adopted. The RBF kernel has two
maximum depth was set from 1 to 20 for all five training important tuning parameters: cost (C) and Gamma. The
samples in this study. The other parameters, such as cross optimal values of C and Gamma are often estimated with
validation folds and min sample count, were set to 10 the exhausted search method. We systematically tested
(default). five different values of C and Gamma to examine the
SVM constructs a hyperplane or a set of hyperplanes in effect of these two key parameters on the performance of
an infinite dimensional space, which can be used for SVM within the object-based approach.
Table 3. Tabulation of training and testing sample objects for each class
NUMBER OF OBJECTS FOR EACH CLASS

INFRASTRUCTURE
BUILDING / ROOF

S (IMEPERVIOUS
URBAN TREE

GRASSLAND
SOIL / SAND

SURFACE)

TOTAL
SHADOW
WATER
OTHER
SAMPLE

TRAINING
TRAINING

TRAINING
TRAINING

TRAINING

TRAINING
TRAINING

TRAINING

TESTING
TESTING

TESTING
TESTING

TESTING
TESTING

TESTING

TESTING

SAMPLE 1 27 12 23 10 17 7 13 6 10 4 7 3 3 1 100 43
SAMPLE 2 53 23 47 20 33 14 27 11 20 9 13 6 7 3 200 86
SAMPLE 3 80 35 70 30 50 21 40 17 30 13 20 9 10 4 300 129
SAMPLE 4 107 46 93 39 67 29 53 23 40 17 27 11 13 6 400 171
SAMPLE 5 133 57 117 50 83 36 67 29 50 21 33 14 17 7 500 214

2.2.3. Phase 3 - Accuracy Assessment

( f − f 21 )
2
Object-based accuracy assessment is a measure of a x 2
= 12 (9)
statistical output to confirm the quality of classification f12 + f 21
results. The method that is most often used to assess
accuracy is based on an error matrix. It utilises appropriate where f12 and f21 indicate the number of ground truth
accuracy measures to compare different classification data samples accurately classified in a set of classification
techniques [39,40]. An error matrix is a cross tabulation of but inaccurately classified in another classifier. These
the classes of the classified imagery and reference values are extracted from the data obtained from
data. It offers a form of site-specific assessment of the the classified image performed by classifiers 1 and 2
correspondence or accuracy degree of the classified image [44].
and the objects in the site [40]. In general, overall Three classes of land cover, including building/roof and
accuracy, producer accuracy, user accuracy and kappa impervious surface (drainage, road), were selected for as-
coefficient are computed from an error matrix [41]. built geometrical assessment. Ground truth data were
McNemar test was performed to examine the role of collected using survey equipment, such as total station for
each classifier. This test identifies a change in the the area. The area of assessment was selected as the area
proportion of the paired data to determine whether the of staff quarters in INSTUN.
statistical differences in classification accuracies are
quantitatively significant [42]. This statistical test uses a
non-parametric approach based on the statistics of a 2 × 2 3. Results and Discussions
matrix [43]. The assessment relies on chi square (x2)
distribution and indicates the statistical differences by 3.1. Results of Preliminary Segmentation
measuring the z score under the null hypothesis that
classification is different. A z score > 1.645 shows the Figure 4(a) to Figure 4(i) show the result of the
confidence level at 95% quantitative significance (p-value preliminary segmentation using the trial and error method.
of 0.05) with one degree of freedom [43]. The McNemar The result shows that the scale parameters of 25, 50, 75,
test represents the statistical difference by measuring a z 100 and 125 results in relevant and acceptable segmentation
score under the null hypothesis that the classifications are against other under-segmented objects for the seven
different. The formula is expressed as follows: selected classes. The result was selected based on the
 Journal of Geosciences and Geomatics 110

criterion of segmented images, that is, an acceptable five best segmentation results with different scale parameters
under-segmented area with a minimum number of were selected for further optimisation segmentation using
over-segmented and under-segmented objects. Therefore, the the Taguchi method.

Figure 4a. (scale: 5) Figure 4d. (scale: 75)

Figure 4e. (scale: 125)

Figure 4b. (scale: 25)

Figure 4c. (scale: 50) Figure 4f. (scale: 125)

111 Journal of Geosciences and Geomatics

Figure 4g. (scale: 150) Figure 4i. (scale: 200)

Figure 4h. (scale: 175) Figure 4j. (scale: 225)

Table 4. Calculation of the plateau of function

INTRASEGMENT

INTRASEGMENT

FUNCTION (POF)
COMPACTNESS

MORAN INDEX
NORMALISED
NORMALISED
VARIANCE, V

PLATEU OF
VARIANCE

VARIANCE
OBJECTS

INDEX, I
MORAN
AREA x
NO. OF
SCALE

SHAPE
TEST

T1 25 0.1 0.1 292,213 284130067.530 2.249 0.616 1.000 0.474 1.474

T2 25 0.3 0.3 287,426 298346842.852 2.361 0.648 0.943 0.392 1.335
T3 25 0.5 0.5 260,712 323533915.878 2.561 0.695 0.841 0.271 1.112
T4 25 0.7 0.7 234,312 357738466.997 2.831 0.717 0.703 0.216 0.920
T5 25 0.9 0.9 218,477 402621039.637 3.187 0.801 0.522 0.000 0.522
T6 50 0.1 0.3 87,752 328136522.871 2.597 0.537 0.823 0.678 1.500
T7 50 0.3 0.5 85,306 348098896.005 2.755 0.590 0.742 0.5411 1.283
T8 50 0.5 0.7 77,416 378344200.146 2.994 0.595 0.620 0.528 1.148
T9 50 0.7 0.9 70,892 414181950.923 3.278 0.627 0.476 0.447 0.923
T10 50 0.9 0.1 34,039 441788694.166 3.497 0.660 0.364 0.362 0.726
T11 75 0.1 0.5 43,599 355906725.097 2.817 0.492 0.711 0.791 1.502
T12 75 0.3 0.7 41,928 381758887.938 3.021 0.539 0.606 0.673 1.279
T13 75 0.5 0.9 38,796 414811875.785 3.283 0.540 0.473 0.671 1.143
T14 75 0.7 0.1 25,934 419955643.806 3.324 0.552 0.452 0.639 1.091
T15 75 0.9 0.3 17,682 481076393.238 3.807 0.572 0.206 0.588 0.794
T16 100 0.1 0.7 26,753 377440063.122 2.987 0.447 0.624 0.907 1.530
T17 100 0.3 0.9 25,778 409356725.481 3.240 0.467 0.495 0.857 1.352
T18 100 0.5 0.1 20,111 412286184.941 3.263 0.513 0.483 0.737 1.221
T19 100 0.7 0.3 15,637 447083102.610 3.538 0.493 0.343 0.790 1.133
T20 100 0.9 0.5 11,949 509986184.541 4.036 0.527 0.089 0.702 0.791
T21 125 0.1 0.9 18,389 396797572.505 3.140 0.411 0.546 1.000 1.546
T22 125 0.3 0.1 16,072 407003727.068 3.221 0.433 0.505 0.943 1.447
T23 125 0.5 0.3 13,728 432107332.699 3.420 0.442 0.403 0.920 1.324
T24 125 0.7 0.5 10,878 470710222.563 3.725 0.430 0.248 0.951 1.199
T25 125 0.9 0.7 9,029 532115563.727 4.211 0.493 0.000 0.791 0.791
 Journal of Geosciences and Geomatics 112

3.2. Results of Optimised Segmentation Using Table 5a. Response table for mean

the Taguchi Method LEVEL SCALE SHAPE COMPACTNESS

1 1.0726 1.5105 1.1919
Table 4 shows the result of each test in which several 2 1.1162 1.3393 1.2171
equations were evaluated to identify the optimised 3 1.162 1.1896 1.1774
parameter for image segmentation. Refer to Table 4, it
4 1.2054 1.053 1.1337
shows that the highest score for POF was the combination
5 1.2613 0.7249 1.0973
of scale, shape and compactness in levels 5, 1 and 5,
respectively, with a score of 1.546. Further interpretation Delta 0.1887 0.7856 0.1198
of the table reveals that a POF value with the highest POF Rank 2 1 3
in a pattern is in agreement with the analysed SN ratio
Table 5b. Response table for SN ratio (larger is better)
result [14]. The hybrid strategy slightly carried the
orthogonal vectors and POF to calculate the SN ratios. LEVEL SCALE SHAPE COMPACTNESS
Figures 5(a) and 5(b) present the main effect plots of 1 0.08617 3.58125 1.25871
the means and SN ratios for multi-resolution segmentation. 2 0.6831 2.52891 1.50894
The response table for mean and SN ratio obtained 3 1.118 1.49106 1.23246
from the analysis of the Taguchi method is presented in 4 1.41904 0.39794 0.85407
Tables 5a and 5b. The result shows that the optimum 5 1.79329 -2.89954 0.24543
combination yielded the highest value of SN ratios and
Delta 1.70712 6.48079 1.26351
means with the associate of level 5 (125) for scale, level 1
Rank 2 1 3
(0.1) for shape and level 2 (0.3) for compactness.

M ain Effects Plot for M eans

Data M eans
SCALE SHAPE COM PACTNESS
1.6

1.5

1.4
M ean of M eans

1.3

1.2

1.1

1.0

0.9

0.8

0.7
25 50 75 100 125 0.1 0.3 0.5 0.7 0.9 0.1 0.3 0.5 0.7 0.9

Figure 5a. Main effect plot (data means)

M ain Effects Plot for SN ratios

Data M eans
SCALE SHAPE COMPACTNESS
4

3
M ean of SN ratios

-1

-2

-3
25 50 75 100 125 0.1 0.3 0.5 0.7 0.9 0.1 0.3 0.5 0.7 0.9

Signal-to-noise: Larger is better

Figure 5b. Main effect plot (SN ratio)

113 Journal of Geosciences and Geomatics

Figure 6a. Building & bare soil Figure 6e. Volleyball court (Impervious Surface)

Figure 6b. Road Figure 6f. Shadow

Figure 6c. Water bodies & grassland Figure 6g. Bare soil & grassland

Figure 6d. Urban tree Figure 6h. Water bodies & Jetty (Impervious Surface)
 Journal of Geosciences and Geomatics 114

Visual judgement confirmed that the optimised C and Gamma were 100,000 and 0.001 for samples 2 and
segmentation parameters yielded the utmost results after 5, respectively, and the result obtained was better than that
applying the Taguchi method, as shown in Figure 6(a) to of the other sample combinations, which affected the
Figure 6(h). Therefore, the merging of the statistical and accuracies of the SVM classification. Most of the data
spatial optimisation processes creates intrinsic sensitivity showed that regardless of the value of C, the overall
to the image pixels and their spatial relationship. The accuracy decreases when Gamma increases to 0.001; the
strength of these properties is utilised to obtain the desired effect is a decrease in OA. The ranges of C and Gamma
quality. factors in all samples indicate that up to 90% accuracy can
be obtained with a Gamma value of 0.1.
3.3. Results of Image Classification Moreover, with a small size of the data sample, a
Gamma parameter of up to 0.0001 may negatively affect
3.3.1. Effects of Parameter Tuning of the Classifiers accuracy, which can only reach up to 50% of OA. On the
contrary, with a sample size of up to 300, OA may
The tuning parameter of the classifier tremendously
consistently be within the minimum of 75% up to 97%
influenced the classification accuracy. The SVM classifier
with increasing sample data. Thus, the effect of tuning
showed high impact and sensitivity to the tuning
parameter is caused by OA, given that a small sample size
parameter, with a variation of up to 60% for each data
(e.g. sample 1 with C and Gamma at 10 and 0.1) results in
sample from the result of minimum to maximum overall
OA of up to 92%. The performance of NB, KNN, RF, DT
accuracy. The variation of KNN showed decreasing
and SVM classifiers with different sample sizes is shown
accuracy with increasing K parameter for all data samples;
in Figure 7 to Figure 8 and Table 6 to Table 11.
it is different from the RF classifier, in which the turning
point is at the maximum depth parameter of 15. When the Table 6. Overall accuracy of NB with the increment in sample data
depth was more than 15, the overall accuracy suddenly
NB CLASSIFIER OA KC
decreased for all data samples. For DT, the trends showed
increasing overall accuracy with the increase in the SAMPLE 1 81.12 0.760
maximum depth parameter for all data samples. However, SAMPLE 2 90.90 0.887
most of the data samples provided the highest result as the SAMPLE 3 91.60 0.896
parameter reached 20.
SAMPLE 4 92.64 0.909
Therefore, the optimal parameter setting for SVM
varied with the data sample size. The optimum values for SAMPLE 5 93.14 0.915

Table 7. Variation of parameter K with proportion to the size of sample data

K=5 K=10 K=15 K=20 K=25
KNN CLASSIFIER
OA KC OA KC OA KC OA KC OA KC
SAMPLE 1 77.62 0.721 69.93 0.623 66.43 0.571 65.03 0.551 61.54 0.506
SAMPLE 2 73.08 0.661 69.23 0.615 65.73 0.562 64.34 0.542 62.24 0.515
SAMPLE 3 72.49 0.662 66.90 0.583 66.20 0.573 63.17 0.536 61.07 0.506
SAMPLE 4 73.38 0.669 68.13 0.596 64.80 0.551 64.27 0.545 62.87 0.524
SAMPLE 5 71.99 0.653 67.79 0.597 65.69 0.564 64.29 0.548 63.45 0.536

Table 8. Variation of depth parameter with proportion to the size of sample data
DEPTH=5 DEPTH=10 DEPTH=15 DEPTH=20 DEPTH=25
RF CLASSIFIER
OA KC OA KC OA KC OA KC OA KC
SAMPLE 1 93.01 0.914 95.10 0.940 95.21 0.942 95.10 0.940 95.10 0.940
SAMPLE 2 76.84 0.717 93.36 0.918 95.80 0.948 95.45 0.944 93.71 0.923
SAMPLE 3 76.22 0.711 89.74 0.874 92.07 0.903 91.38 0.894 91.38 0.894
SAMPLE 4 76.53 0.713 88.79 0.863 92.29 0.906 91.24 0.893 91.24 0.893
SAMPLE 5 74.79 0.694 87.39 0.846 89.08 0.867 88.80 0.863 89.36 0.870

Table 9. Variation of parameter depth with proportion to the size of sample data
DEPTH=5 DEPTH=10 DEPTH=15 DEPTH=20 DEPTH=25
DT CLASSIFIER
OA KC OA KC OA KC OA KC OA KC
SAMPLE 1 81.82 0.776 85.31 0.821 85.31 0.821 86.01 0.830 85.31 0.821
SAMPLE 2 73.08 0.670 83.92 0.805 83.22 0.796 84.27 0.809 83.22 0.796
SAMPLE 3 68.53 0.618 75.99 0.711 75.99 0.711 76.13 0.717 75.76 0.709
SAMPLE 4 70.75 0.644 83.19 0.796 82.66 0.790 82.66 0.790 82.66 0.790
SAMPLE 5 71.57 0.646 77.87 0.731 79.27 0.746 79.13 0.745 79.31 0.749
115 Journal of Geosciences and Geomatics

Table 10. Variation of the combination of parameter C and Gamma with proportion to the size of sample data
SVM CLASSIFIER SAMPLE 1 SAMPLE 2 SAMPLE 3 SAMPLE 4 SAMPLE 5
C GAMMA OA KC OA KC OA KC OA KC OA KC
1 78.32 0.724 79.02 0.735 78.79 0.731 78.46 0.727 77.17 0.714
0.1 92.31 0.905 89.51 0.870 87.88 0.850 88.97 0.863 77.17 0.714
10 0.01 81.12 0.763 74.83 0.681 75.99 0.698 76.01 0.698 76.19 0.702
0.001 54.55 0.390 51.75 0.354 55.01 0.403 56.57 0.424 55.18 0.405
0.0001 27.27 0.002 26.57 0.001 26.81 0.001 32.22 0.077 34.31 0.108
1 78.32 0.724 79.02 0.735 79.02 0.734 78.46 0.727 78.71 0.730
0.1 93.71 0.922 92.31 0.905 92.77 0.911 93.17 0.916 93.13 0.916
100 0.01 90.91 0.888 89.16 0.866 88.34 0.856 90.89 0.887 92.58 0.908
0.001 83.22 0.790 75.17 0.686 75.52 0.692 76.88 0.710 77.73 0.722
0.0001 53.15 0.371 51.75 0.354 55.71 0.413 56.57 0.425 55.00 0.402
1 78.32 0.724 79.02 0.735 79.02 0.734 78.46 0.727 78.71 0.730
0.1 93.71 0.922 92.31 0.905 93.01 0.914 92.99 0.913 93.00 0.914
1000 0.01 93.01 0.914 95.10 0.940 93.47 0.920 96.15 0.953 95.94 0.950
0.001 90.91 0.888 89.86 0.875 90.21 0.879 90.01 0.876 90.76 0.886
0.0001 82.52 0.781 76.57 0.704 77.16 0.714 76.36 0.703 77.17 0.714
1 78.32 0.724 79.02 0.735 79.02 0.734 78.46 0.727 78.71 0.730
0.1 93.82 0.931 92.66 0.910 93.01 0.914 92.99 0.913 93.28 0.917
10000 0.01 93.01 0.914 94.76 0.935 94.41 0.931 96.82 0.958 96.78 0.960
0.001 92.31 0.905 95.10 0.940 95.80 0.948 95.80 0.948 96.80 0.961
0.0001 90.91 0.888 88.11 0.853 88.58 0.859 88.62 0.859 89.64 0.872
1 78.32 0.724 79.02 0.735 79.02 0.734 78.46 0.727 78.71 0.730
0.1 93.71 0.922 92.66 0.910 93.01 0.914 92.99 0.913 93.00 0.914
100000 0.01 93.01 0.914 95.10 0.940 94.64 0.934 96.15 0.953 96.50 0.957
0.001 92.31 0.905 96.15 0.953 95.57 0.946 96.15 0.953 97.20 0.966
0.0001 92.31 0.905 95.45 0.944 95.10 0.940 93.35 0.918 94.82 0.936

3.3.2. Effects of Varying the Number of Selected The size of training and testing data affected the
Samples classification accuracy of certain classifiers, such as NB
The result of the analysis showed that SVM obtained and DT, compared with RF and KNN, which are less sensitive
the highest accuracy among the five classifiers in terms of to the increase in sample data. SVM and RF obtained a
C and Gamma. The minimum result for the overall consistent overall accuracy with the most outstanding
accuracy and kappa coefficient of SVM for all samples result (more than 90%) compared with the other classifiers.
was 93.82% and 0.931, respectively, with the total average The KNN classifier obtained the highest score of 77.62%
of five samples at 95.96% and 0.951. In addition, the and 0.721 for accuracy and kappa values, respectively.
overall accuracy of SVM was maintained with an Hence, increasing the number of data samples did not
accuracy of more than 90% each. The minimum accuracy affect the increase in the overall accuracy of classification.
of the SVM classifier was better than the highest overall The result shows that the variation of the size of
accuracy and kappa coefficient for NB (93.14% and training samples from 100 to 500 (sample 1 to 5)
0.915), DT (86.01% and 0.830) and KNN (77.62% and increased the accuracy of NB and SVM by 12.02% and
0.721). Sample 2 of RF showed the highest overall 3.38%, respectively. On the contrary, the accuracy of DT,
accuracy and kappa coefficient of 95.80% and 0.948, RF and KNN decreased to 6.74%, 5.85% and 5.63%,
respectively. Sample 5 of SVM obtained the highest result. respectively. The NB classifier was the most sensitive to
The result is contrary to that of Sample 5 of RF, which the variation of sample size because the parametric
obtained the lowest overall accuracy and kappa coefficient classifier consumes the training samples to estimate the
with the maximum size of sample data. The overall parameter value for data allocation. Hence, with the
accuracy for NB was consistent in Samples 2 to 5, with an increasing number of training samples, a highly accurate
accuracy of more than 90% each. However, with a small parameter estimation can be achieved. SVM is the least
number of training and testing data samples, the accuracy sensitive to sample sizes because it requires support
decreased to 81.12%. The average accuracy and kappa vectors rather than using all training samples to create
coefficient of the DT classifier were 81.77% and 0.780, a separating hyperplane. Table 11 shows that the
respectively, and Sample 1 obtained the best result. classification accuracy of the three classifiers shifted and
Graphs in Figure 7 and Figure 8 show that KNN obtained became inconsistent when Sample 3 with more than 300
the lowest accuracy and kappa values for all data sample samples was selected with a variation of DT (-8.74%), RF
sizes compared with the other classifiers. (-3.73%) and KNN (-0.59%). This result revealed that
 Journal of Geosciences and Geomatics 116

Sample 3 is the turning point among these classifiers. The additional sample size disrupted the accuracy.
Table 11. The best accuracy for each classifier with the variation of the size of sample data

NB KNN RF DT SVM
CLASSIFIER / SAMPLE
OA KC OA KC OA KC OA KC OA KC
SAMPLE 1 81.12 0.760 77.62 0.721 95.21 0.942 86.01 0.830 93.82 0.931
SAMPLE 2 90.90 0.887 73.08 0.661 95.80 0.948 84.27 0.809 96.15 0.953
SAMPLE 3 91.60 0.896 72.49 0.662 92.07 0.903 76.13 0.717 95.80 0.948
SAMPLE 4 92.64 0.909 73.38 0.669 92.29 0.906 83.19 0.796 96.82 0.958
SAMPLE 5 93.14 0.915 71.99 0.653 89.36 0.870 79.27 0.746 97.20 0.966
AVERAGE 89.88 0.873 73.71 0.673 92.95 0.914 81.77 0.780 95.96 0.951

Figure 7. Result of overall accuracies for the best of the five classifiers with increasing sample size

CLASSIFIERS KC PERFORMANCE
1.000

0.900

0.800

0.700

0.600
NB
KC

0.500 KNN

RF
0.400
DT
0.300 SVM

0.200

0.100

0.000
0 1 2 3 4 5 6

SAMPLE

Figure 8. Kappa coefficient (KIA) of the five classifiers with increasing sample size
117 Journal of Geosciences and Geomatics

Figure 9a. Orthomosaic Figure 9c. RF classification

Figure 9b. SVM classification Figure 9d. NB classification

 Journal of Geosciences and Geomatics 118

Figure 9e. DT classification Figure 9f. KNN classification

Table 12. The best classification accuracy for each classifier

CLASSIFIER SAMPLE DATA PARAMETER VALUE OVERALL ACCURACY KIA

NB 5 N/A N/A 93.14 0.915
KNN 1 K NEIGHBORS K=5 77.62 0.721
RF 2 MAXIMUM DEPTH D=15 95.80 0.948
DT 1 MAXIMUM DEPTH D=20 86.01 0.830
SVM 5 GAMMA, C GAMMA = 0.001, C=100000 97.20 0.966

Table 13. Results of the McNemar test for each classifier

CLASSIFIER 1 CLASSIFIER 2 X2 P value SIGNIFICANT
SVM RF 13.39535 0.0025 SIGNIFICANT
SVM NB 9.63333 0.00191 SIGNIFICANT
SVM DT 62.02532 0 SIGNIFICANT
SVM KNN 51.8481 0 SIGNIFICANT
RF NB 0.70588 0.4081 NOT SIGNIFICANT
RF DT 31.64063 0 SIGNIFICANT
RF KNN 27.16071 0 SIGNIFICANT
NB DT 29.71429 0 SIGNIFICANT
NB KNN 31.58209 0 SIGNIFICANT
DT KNN 0.28409 0.59403 NOT SIGNIFICANT

Figures 9(a) to 9(f) present an orthomosaic of the study classifier when the training samples are sufficiently large.
area and the result of the best classification for each RF also shows the potential to obtain high accuracies as
classifier with the parameter indicated in Table 12. These those of SVM with sufficient parameter setting.
findings present an important implication in the selection The McNemar test indicated that the SVM classifier
of appropriate classifiers. SVM outperformed the other was highly significant compared with the other classifiers.
classifiers and was the best classifier for land cover Hence, SVM outperformed the other classifiers. The
classification. The SVM classifier provided good overall significant results of the McNemar test are presented in
classification for all sample data. NB can be an alternative Table 13.
119 Journal of Geosciences and Geomatics

3.4. Error Assessment of Geometrical assessment were building, road and drainage.
Information The results show that the difference of the ground truth
data with the image classification was less than 1 metre.
In conjunction with best classification result of the The geometry of the as-built feature is acceptable because
SVM classifier, the image underwent geometrical assessment the error is less than 1 m and is justified for mapping
with the dimensional features evaluated. The three types purposes as Table 14. The final topographic map is shown
of civil infrastructures selected for the geometrical in Figure 10.

Table 14. Geometrical assessment of as-built features

GROUND TRUTH (Acquired from Ground Truth Survey) ERROR ASSESSMENT

CLASSES
& OBIA (Extraction from UAV data) = Ground Truth (GT) - OBIA

CT

CLASSES GT (m) OBIA (m) ERROR +/-

B1 (a) 10.809 10.94 -0.131
Building 1
BUILDING, B1 (b) 5.513 5.613 -0.100
B1 B1 (c) 8.982 8.875 0.107
B1 (d) 9.989 9.915 0.074

OBIA

GT

CLASSES GT (m) OBIA (m) ERROR +/-

B2 (a) 8.483 8.397 0.086
Building 2 BUILDING, B2 (b) 13.822 13.939 -0.117
B2 B2 (c) 14.417 14.311 0.106
B2 (d) 5.593 5.744 -0.151

OBIA
 Journal of Geosciences and Geomatics 120

GROUND TRUTH (Acquired from Ground Truth Survey) ERROR ASSESSMENT

CLASSES
& OBIA (Extraction from UAV data) = Ground Truth (GT) - OBIA
CT

CLASSES GT (m) OBIA (m) ERROR +/-

D1 (a) 1.398 1.526 -0.128
Drainage 1 DRAINAGE,
D1 (b) 1.599 1.507 0.092
OBIA D1
D1 (c) 10.462 10.606 -0.144

GT

CLASSES GT (m) OBIA (m) ERROR +/-

D2 (a) 12.865 12.756 0.109
Drainage 2 DRAINAGE, D2 (b) 1.313 1.517 -0.204
D2 D2 (c) 1.338 1.255 0.083
D2 (d) 3.308 3.511 -0.203

OBIA

GT

CLASSES GT (m) OBIA (m) ERROR +/-

R1 (a) 14.993 14.880 0.113
Road 1 ROAD, R1 (b) 12.425 12.307 0.118
R1 R1 (c) 11.238 11.114 0.124
R1 (d) 5.533 5.638 -0.105

OBIA
121 Journal of Geosciences and Geomatics

GROUND TRUTH (Acquired from Ground Truth Survey) ERROR ASSESSMENT

CLASSES
& OBIA (Extraction from UAV data) = Ground Truth (GT) - OBIA

GROUND TRUTH
ERROR ASSESSMENT
CLASSES (Acquired from Ground Truth Survey) & OBIA
= Ground Truth (GT) - OBIA
(Extraction from UAV data)

GT

CLASSES GT (m) OBIA (m) ERROR +/-

R2 (a) 5.597 5.482 0.115
Road 2 ROAD, R2 (b) 10.501 10.59 -0.089
R2 R2 (c) 5.595 5.711 -0.116
R2 (d) 29.234 29.351 -0.117

OBIA

4. Conclusion
We have evaluated and compared the performance of
five machine learning classifiers in classifying high-
resolution images by implementing an OBIA procedure.
The SVM classifier obtained the highest results among all
classifiers. The classification was affected by the accuracy
of the tuning parameters for each classifier when different
sizes of training and testing samples were used.
The extraction of as-built information was examined
through a geometrical assessment with the output of the
data extraction. The tolerance of as-built information
(building, drainage and road) through the ground truth
data using survey equipment (total station) with UAV
OBIA data obtained an error of less than 1 metre. The
combination of UAV and OBIA can provide a rapid and
efficient approach for map updating, especially in rapidly
changing urban areas. This technique can potentially
replace current procedures that utilise piloted aircraft and
high-resolution satellite data (more expensive and time
consuming). Hence, the results of this study provide an
additional insight into the use of OBIA for UAV optical
imagery information extraction.

Acknowledgements
The authors would like to thank the National Land and
Survey Institute for providing and sharing UAV data to
realise this research. Gratitude is also extended to the staff
Figure 10. Final topographic map based on the best accuracy of the
SVM classifier
of Photogrammetry Section, Division of Topography
Mapping, Department of Survey and Mapping Malaysia
 Journal of Geosciences and Geomatics 122

(JUPEM) for their contributions, ideas, comments and [17] Majeed, Z. A., Saip, S. N., & Ng, E. G. (2017). Towards
criticisms on the current issues related to the scope of the augmented topographic map: Integration of digital photograph
captured from MAV and UAV platform. In FIG Working Week
study. UPM is also acknowledged for the provision of 2017 Presentation Paper (pp. 1-15).
financial support via the GRF scheme. Comment from the [18] Ma, L., Cheng, L., Li, M., Liu, Y., & Ma, X. (2015). Training set
anonymous reviewers are highly appreciated. size , scale , and features in Geographic Object-Based Image
Analysis of very high resolution unmanned aerial vehicle imagery.
ISPRS Journal of Photogrammetry and Remote Sensing, 102,
14-27.
References [19] Qian, Y., Zhou, W., Yan, J., Li, W., & Han, L. (2015). Comparing
Machine Learning Classifiers for Object-Based Land Cover
[1] Ahmad, A., Hashim, K. A., & Samad, A. M. (2010). Aerial Classification Using Very High Resolution Imagery. Remote
Mapping using High Resolution Digital Camera and Unmanned Sensing, 7, 153-168.
Aerial Vehicle for Geographical Information System. 2010 6th [20] Sharma, J. B., & Hulsey, D. (2014). Integrating the UAS in
International Colloquium on Signal Processing & Its Applications Undergradute Teaching and Research - Oppurtunity and
(CSPA), 201-206. Challanges at The University of Georgia. In The International
[2] Hardin, P. J., & Hardin, T. J. (2010). Small-scale remotely piloted Archives of the Photogrammetry, Remote Sensing and Spatial
vehicles in environmental research. Geography Compass, 4(9), Information Sciences. ISPRS Technical Commission I Symposium
1297-1311. (Vol. XL, pp. 17-20).
[3] Laliberte, A. S., & Rango, A. (2009). Texture and Scale in Object- [21] Grigillo, D., & Kanjir, U. (2012). Urban object extraction from
Based Analysis of Subdecimeter Resolution Unmanned Aerial digital surface model and digital aerial images. In ISPRS Annals of
Vehicle (UAV) Imagery. IEEE Transactions on Geoscience and the Photogrammetry, Remote Sensing and Spatial Information
Remote Sensing, 47(3), 761-770. Sciences (Vol. 1-3, pp. 215-220).
[4] Nex, F., & Remondino, F. (2014). UAV for 3D mapping [22] Murcko, J. (2017). Object-based classification for estimation of
applications: A review. Applied Geomatics, 6(1), 1-15. built-up density within urban environment.
[5] Ahmad, A. (2016). The Direction of UAV Technology in Malaysia [23] Suzuki, K., Liu, W., Estrada, M., & Yamazaki, F. (2013). Object-
& The Principle of Photogrammetry for UAV Mapping. based building extraction in Tacna, Peru using worldview-2
PowerPoint presentation at the courses of UAV Technology and images. In Proceedings of ACRS 2013 (pp. 1159-1166).
Image Processing, The National and Survey Institute (INSTUN), [24] Tomljenovic, I., Tiede, D., & Blaschke, T. (2016). A building
Perak, Malaysia. extraction approach for Airborne Laser Scanner data utilizing the
[6] Mohammadi, M., Hahn, M., & J, E. (2011). Road Classification Object Based Image Analysis paradigm. International Journal of
and Condition Investigation Using Hyperspectral Imagery. In Applied Earth Observations and Geoinformation, 52, 137-148.
Applied Geoinformatics for Society and Environment. Jomo [25] Kavzoglu, T., & Yildiz, M. (2014). Parameter-Based Performance
Kenyatta University of Agriculture and Technology Stuttgart Analysis of Object-Based Image Analysis Using Aerial and
University of Applied Sciences. Quikbird-2 Images. ISPRS Annals of the Photogrammetry, Remote
[7] Crommelinck, S., Bennett, R., Gerke, M., Nex, F., Yang, M. Y., & Sensing and Spatial Information Sciences, Volume II-7, 2014,
Vosselman, G. (2016). Review of automatic feature extraction II(October), 31-37.
from high-resolution optical sensor data for UAV-based cadastral [26] Baatz, M., & Schäpe, A. (2000). Multiresolution segmentation: An
mapping. Remote Sensing, 8(689). optimization approach for high quality multi-scale image
[8] Jazayeri, I., Rajabifard, A., & Kalantari, M. (2014). A geometric segmentation. Proceedings of Angewandte Geographische
and semantic evaluation of 3D data sourcing methods for land and Informationsverarbeitung XII, 12-23.
property information. Land Use Policy, 36, 219-230. [27] Martha, T. R., Kerle, N., Van Westen, C. J., Jetten, V., & Kumar,
[9] Blaschke, T., Lang, S., Lorup, E., Strobl, J., & Zeil, P. (2000). K. V. (2011). Segment optimization and data-driven thresholding
Object-Oriented Image Processing in an Integrated GIS / Remote for knowledge-based landslide detection by object-based image
Sensing Environment and Perspectives for Environmental analysis. IEEE Transactions on Geoscience and Remote Sensing,
Applications. Environmental Information for Planning, Politics 49(12 PART 1), 4928-4943.
and the Public, (1995), 555-570. [28] Li, C., & Shao, G. (2012). Object-oriented classification of land
[10] Gibril, M. B. A., Shafri, H. Z. M., & Hamedianfar, A. (2017). use / cover using digital aerial orthophotography. International
New semi-automated mapping of asbestos cement roofs using Journal of Remote Sensing, 33, 922-938.
rule-based object-based image analysis and Taguchi optimization [29] Lowe, S. H., & Guo, X. (2011). Detecting an Optimal Scale
technique from WorldView-2 images. International Journal of Parameter in Object-Oriented Classification. IEEE Journal of
Remote Sensing, 38(2), 467-491. Selected Topics in Applied Earth Observations and Remote
[11] Ma, L., Li, M., Ma, X., Cheng, L., Du, P., & Liu, Y. (2017). A Sensing, 4(4), 890-895.
review of supervised object-based land-cover image classification. [30] Pu, R., Landry, S., & Yu, Q. (2011). Object-based urban detailed
ISPRS Journal of Photogrammetry and Remote Sensing, 130, land cover classification with high spatial resolution IKONOS
277-293. imagery, 32(12), 3285-3308.
[12] Tzotsos, A., Karantzalos, K., & Argialas, D. (2010). Object-based [31] Chou, C.-S., Ho, C.-Y., & Huang, C.-I. (2009). The optimum
image analysis through nonlinear scale-space filtering. ISPRS conditions for comminution of magnetic particles driven by a
Journal of Photogrammetry and Remote Sensing, 66(1), 2-16. rotating magnetic field using the Taguchi method. Advanced
[13] Rao, R. S., Kumar, C. G., Prakasham, R. S., & Hobbs, P. J. (2008). Powder Technology, 20(1), 55-61.
The Taguchi methodology as a statistical tool for biotechnological [32] Aggarwal, A., Singh, H., Kumar, P., & Singh, M. (2008).
applications: A critical appraisal. Biotechnology Journal, 3(4), Optimizing power consumption for CNC turned parts using
510-523. response surface methodology and Taguchi’s technique-A
[14] Idrees, M. O., & Pradhan, B. (2016). Hybrid Taguchi-Objective comparative analysis. Journal of Materials Processing Technology,
Function optimization approach for automatic cave bird detection 200(1-3), 373-384.
from terrestrial laser scanning intensity image. International [33] Pradhan, B., Jebur, M. N., Shafi, H. Z. M., & Tehrany, M. S.
Journal of Speleology, 45(3), 289-301. (2015). Data Fusion Technique Using Wavelet Transform and
[15] Moosavi, V., Talebi, A., & Shirmohammadi, B. (2013). Producing Taguchi Methods for Automatic Landslide Detection From
a landslide inventory map using pixel-based and object-oriented Airborne Laser Scanning Data and QuickBird Satellite Imagery.
approaches optimized by Taguchi method Vahid. Geomorphology. IEEE Transactions on Geoscience and Remote Sensing, 1-13.
[16] Sameen, M. I., & Pradhan, B. (2017). A Two-Stage Optimization [34] Espindola, G. M., Camara, G., Reis, I. A., Bins, L. S., &
Strategy for Fuzzy Object-Based Analysis Using Airborne LiDAR Monteiro, A. M. (2006). Parameter selection for region-growing
and High-Resolution Orthophotos for Urban Road Extraction, image segmentation algorithms using spatial autocorrelation.
2017. International Journal of Remote Sensing, 27(14), 3035-3040.
123 Journal of Geosciences and Geomatics

[35] Gao, Y., Kerle, N., Mas, J. F., Navarrete, A., & Niemeyer, I. [43] Foody, G. M. (2004). Thematic map comparison: evaluating the
(2007). Optimized Image Segmentation and Its Effect on statistical significance of differences in classification accuracy.
Classification Accuracy. 5th International Symposium - Spatial Photogrammetric Engineering & Remote Sensing, 70(5), 627-633.
Data Quality, 4p. Retrieved from [44] Leeuw, J. D., Jia, H., Yang, L., Schmidt, K., & Skidmore, A. K.
http://www.itc.nl/ISSDQ2007/proceedings/postersession/gao (2006). Comparing accuracy assessment to infer superiority of
kerleet al.pdf. image classification methods. International Journal of Remote
[36] Fotheringham, A. S., Brunsdon, C., & Charlton, M. (2000). Sensing, 27(1), 223-232.
Quantitative Geography: Perspectives on Spatial Data Analysis,. [45] Goebel, K., & Saha, B. (2015). Handbook of Unmanned Aerial
Cleveland State University. Vehicles. Springer Reference. Springer Dordrecht Heidelberg New
[37] Wieland, M., & Pittore, M. (2014). Performance evaluation of York London.
machine learning algorithms for urban pattern recognition from [46] Husran, Z (2016, November 8). The use of UAV Technology in
multi-spectral satellite images. Remote Sensing, 6(4), 2912-2939. Planning and Security Monitoring. PowerPoint presentation at the
[38] Trimble. (2014). eCognition ® Developer. User Guide. courses of UAV Technology and Image Processing, The National
[39] Nichol, J., & Wong, M. S. (2008). Habitat Mapping in Rugged and Survey Institute (INSTUN), Perak, Malaysia.
Terrain Using Multispectral Ikonos Images. Photogrammetric [47] International Civil Aviation Organization (ICAO) (Ed.). (2011).
Engineering & Remote Sensing, 74(11), 1325-1334. Unmanned Aircraft Systems (UAS). University Street, Montréal,
[40] Foody, G. M. (2002). Status of land cover classification accuracy Quebec, Canada.
assessment. Remote Sensing of Environment, 80(1), 185-201. [48] JUPEM (2010). JUPEM: A pictorial journey 1885-2010
[41] Congalton, R. G., & Green, K. (1993). A Practical Look at the (Commemorating the 125th Anniversary). Kuala Lumpur,
Sources of Confusion in Error Matrix Generation. American Malaysia
Society for Photogrammetry and Remote Sensing, 59(5), 641-644. [49] Rabab, M.Z.M (2012). UAV capabilities for the purpose of Data
[42] Hamedianfar, A., Shafri, H. Z. M., Mansor, S., & Ahmad, N. Acquisition for geospatial defence data. Berita Ukur July –
(2014). Improving detailed rule-based feature extraction of urban December 2012. Malaysia.
areas from WorldView-2 image and lidar data. International [50] Wang, T. Y., & Huang, C. Y. (2007). Improving forecasting
Journal of Remote Sensing, 35(5), 1876-1899. performance by employing the Taguchi method. European
Journal of Operational Research, 176(2), 1052-1065.