> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) < 1

Clustering green open space using UAV (Unmanned Aerial Vehicle) with
CNN (Convolutional Neural Network)
M. Y. Fikri(1), M. J. Al Firdaus (2), K. Azzarkhiyah (3), M. Syai’in (4)

II. METHODOLOGY
Abstract—The latest in unmanned aerial vehicles (UAVs) and The purpose of this research is to classify between a trees
associated sensing systems make these increasingly attractive and no tree using Convolution Neural Network (CNN) so that
platforms to the remote sensing community. A large number of
the drone can capture an image automatically and calculate
spatial details contained in these images opens the door for
advanced monitoring applications. In this paper, we use this
percentage of green open space. This research is applied to
cost-effective and attractive technology for the automatic count the number of trees and can help know enough or not
detection of green open spaces. Given a UAV image of trees green open space in a given area of yield is in the form of a
acquired, then, we analyze these Convolutional Neural Networks percentage of green land to the total land area in the region.
(CNN) points of the prior classifier trained on a set of trees and
no trees points. As output, CNN will mark each detected tree by
super pixel. Then, in order to capture the shape of each tree, we
propose to merge this pixel-level segmentation with a method
based active contour on the Color threshold. Finally, we further
analyze the texture of regions with pixel-level segmentation and
use summing pixel to distinguish trees from other vegetation.
Experimental results obtained in UAV images from extensive
calculations using the program that has been made and the
existing provisions get a result of error of 7.256% on the first
trial, the second experiment is 5.156%, and the third
experiment is 3.126%.

Keywords—Unmanned Aerial Vehicles, Convolutional Neural

Network, super pixel Fig.1 block diagram for training CNN

I. INTRODUCTION
Current technological developments are running very fast
in almost all aspects. One of them is technology utilization in
the geographical mapping of a region. Geographical
information is one of the important and very necessary
aspects for some parties in certain circumstances, such as
mapping land functions from an area.
A. RGB
RGB color space is the main three colors (the three
primary colors) which consists of three color channels, red,
green, and blue. Each color channel is limited to a value of 0
- 255 or in other words there can be 256 levels. If the color
channel is combined then there are 256 x 256 x256 or
16,777,216 RGB color combinations produced.

Mapping of land from a certain area can be done using

digital imagery with a shooting technique called
photogrammetry. Photogrammetry is an art, science, and
technology to obtain reliable information about physical
objects and the environment through the process of recording,
measuring, and interpreting photographic images and
recorded patterns of electromagnetic energy [] One of the
problems that can be solved by photogrammetry is in the field
of forestry, namely efforts to monitor forest land such as Fig.2 contoh gambar RGB
counting the number of trees and calculating the percentage B. Grayscale
of green land in an area by utilizing digital imagery. A grayscale image is an image that only has a gray level
color. In this process the grayscale image is used for the
Iki tambahi pid cek ketok akeh information given to each pixel compared to the color image.
coiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiii Gray color can be said as a color with one color channel,
iiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiii which is an 8 bit color channel.
iiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiii
iiiiiiiiiiiiiiiiiiiiiii


> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) < 2

Fig.3 proses mengubah RGB menjadi grayscale

C. Convolution
Convolution is a filtering image process that is often
carried out in image processing. In this study, image
convolution uses a 3x3 size convolution carnelian with a gray
image measuring 640x480. In the image convolution stage,
the author convolutes using the kernel twice to clarify the
image so that it will get a more detailed image. The first
convolution process uses gray images with a size of 640x480
and 32 kernels which produce a new array measuring Fig.5 proses maxpooling
638x478. The second convolution of 319x239 and 64 kernel
gray images that produced a new array measuring 317x237. E. Training Neural Network
On Supervised Learning using a Neural Network, it
generally consists of two stages, namely training and
evaluation. But there are also additional stages, namely
testing, but this stage is not mandatory. In the training stage,
the author will update the weight and bias on each neuron
continuously so that the output produced is in line with
expectations. In each iteration an evaluation process will be
carried out which serves to determine the time the training
process has been completed (stopping point).
In the next part, we will discuss how to process training on
the neural network. But it will be explained in general. The
training process consists of 2 stages: Forward pass or forward
propagation is a process whereby we carry data at the input
Fig.4 proses konvolusi
passing through each neuron in the hidden layer to the output
layer, which will then calculate the error. And the error we get
D. Maxpooling on the forward pass will be used to update each weight and
After passing through the convolution process, the next bias it with a particular learning rate or can use an auto
stage is RELU layer will implement the activation function learning rate.
elements, such as max (0, x) Thresholding zero [9]. So, if the
pixel value is less than zero then it will be switched equal to
zero. Subsampling layer is the process of reducing the size of
an image data. In image processing, subsampling also aims to
improve the position invariance of features. In most of CNN,
subsampling method used is the max pooling. Max pooling
output of the convolution layer divides into several smaller
grids and taking the maximum value of each grid to compose
the image matrix that has been reduced as shown in Fig. 3.
The results of this process can be seen on the set of the grid
on the right. The process will ensure the same features found
even though the object image translated (shift).

Fig.6 contoh struktur training Neural Network yang

digunakan

F. Hardware
Fig.7 Shows some of hardware that is used in this
Maxpooling + Relu research. PC is used as a central processing unit where almost
5x5 all of tasks are processed. The tasks here such as displaying,
capturing, classifying, filtering, and calculate percentage of
green open space. The results of the classification will trigger
> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) <
programs that have been made to capture images through
drones with action cam that has been placed on the drone. To
capture the signal from the camera a dual band Wi-Fi
extender is needed for a more extended range.
Fig.7 hardware yang digunakan dalam penelitian ini
III. TESTING AND DATA ANALYSIS
Data testing uses the CNN method to classify and color
thresholding to calculate the percentage of green open areas
in an area.
Fig.8 diagram blok kerja sistem
A. Video Streaming
What is meant by video streaming in this paper is to
connect between the action of the camera and the laptop using
the dual band 5ghz wifi extender. Then the laptop will get an
ip from the action cam that will be accessed using Visual
Studio 2017.
Fig.9 interface yang digunakan untuk video streaming
B. CNN
This paper is focused on getting an automatic system for
tree classification, whether or not there is. In addition, the
main purpose of this study is to obtain the classification
results of the CNN process. Then it will trigger to capture the
image automatically which will be processed later.
C. Noise Reduction
Since impulse noises generated by sensor or
communication errors usually corrupt images, noise reduction
3
is essential for improving the results of subsequent processing
such as image segmentation and morphological operations[].
In this paper, median filter (a widely used nonlinear digital
filtering technique) is adopted to eliminate noise in images
due to its simplicity and capability of preserving image edges
while removing noise [].
Fig.10 proses noise reduction untuk gambar yang telah
tercapture
D. Green Open Space Segmentation
Segmentation is an important step for Percentage
Calculation. Its main objective is to differentiate green open
space pixels from background pixels. Thresholding is a
frequently adopted technique to segment the green open space
regions in images, while Otsu method is one of the widely
adopted thresholding approach for image segmentation.
Fig.11 hasil proses threshold gambar
E. Morphological Operations
Although noise reduction, color model conversion, and
green open space segmentation have been applied, there still
exist some small irrelative objects, which may affect the green
open space confirmation. In this paper, the solution to this is
to employ the mathematic morphological operations to
remove the small objects which are not target objects in the
percentage calculation.
Mathematic morphological operations contain a series of
operators such as dilation closing, which is capable of
effectively removing small irrelative objects in the
thresholding images. This paper applies dilation, since
erosion can’t get rid of pixels on the object boundaries while
dilation can add pixels.
The erosion operation E and dilation operation D between
image set I and morphological element C can be separately
illustrated as follows:
(1)
(2)
Where symbol “ ”denotes the erosion operation and
symbol “ ”denotes dilation operation. (i,j) denote the
coordinates of pixel in image I.
> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) <
F. Percentage Calculation
Setelah gambar diolah menggunakan morphological
operations maka step selanjutnya adalah mencari berapa
persen luasan green openspace yang ada pada suatu gambar.
Untuk mencari luasan berapa persennya. Penulis
menggunakan summing pixel untuk mengetahui jumlah area
terbuka hijau yang ada. Setelah itu penulis membagi hasil
penjumlahan area terbuka hijau menggunakan luasan pixel
yang ada. Untuk hasil gambar dapat dilihat pada fig.12.
Fig.12 hasil akhir dari proses perhitungan presentase green
openspace
IV. CONCLUSION
Convolution neural network dapat memprediksi dengan
sukses ada atau tidaknya pohon dengan ketepatan hingga
90.35 %. Dengan menggunakan proses konvolusi sebanyak 2
kali. Pada proses konvolusi pertama menggunakan 32 kernel
sedangkan untuk proses konvolusi kedua menggunakan
sebanyak 64 kernel.
Untuk perhitungan ketepatan mendeteksi presentase
jumlah pohon dalam satu gambar mempunyai ketepatan
hampir 90% apabila dibandingkan dengan hitung manual
yang telah ada.
REFERENCES
[1] G. O. Young, “Synthetic structure of industrial plastics (Book style with
paper title and editor),” in Plastics, 2nd ed. vol. 3, J. Peters, Ed. New
York: McGraw-Hill, 1964, pp. 15–64.
[2] W.-K. Chen, Linear Networks and Systems (Book style). Belmont, CA:
Wadsworth, 1993, pp. 123–135.
[3] H. Poor, An Introduction to Signal Detection and Estimation. New York:
Springer-Verlag, 1985, ch. 4.
[4] B. Smith, “An approach to graphs of linear forms (Unpublished work
style),” unpublished.
[5] E. H. Miller, “A note on reflector arrays (Periodical style—Accepted for
publication),” IEEE Trans. Antennas Propagat., to be published.
[6] J. Wang, “Fundamentals of erbium-doped fiber amplifiers arrays
(Periodical style—Submitted for publication),” IEEE J. Quantum
Electron., submitted for publication.
[7] C. J. Kaufman, Rocky Mountain Research Lab., Boulder, CO, private
communication, May 1995.
[8] Y. Yorozu, M. Hirano, K. Oka, and Y. Tagawa, “Electron spectroscopy
studies on magneto-optical media and plastic substrate interfaces
(Translation Journals style),” IEEE Transl. J. Magn.Jpn., vol. 2, Aug.
1987, pp. 740–741 [Dig. 9th Annu. Conf. Magnetics Japan, 1982, p. 301].
[9] M. Young, The Techincal Writers Handbook. Mill Valley, CA: University
Science, 1989.
4
[10] J. U. Duncombe, “Infrared navigation—Part I: An assessment of feasibility
(Periodical style),” IEEE Trans. Electron Devices, vol. ED-11, pp. 34–39,
Jan. 1959.
[11] S. Chen, B. Mulgrew, and P. M. Grant, “A clustering technique for digital
communications channel equalization using radial basis function
networks,” IEEE Trans. Neural Networks, vol. 4, pp. 570–578, Jul. 1993.
[12] R. W. Lucky, “Automatic equalization for digital communication,” Bell
Syst. Tech. J., vol. 44, no. 4, pp. 547–588, Apr. 1965.
[13] S. P. Bingulac, “On the compatibility of adaptive controllers (Published
Conference Proceedings style),” in Proc. 4th Annu. Allerton Conf. Circuits
and Systems Theory, New York, 1994, pp. 8–16.
[14] G. R. Faulhaber, “Design of service systems with priority reservation,” in
Conf. Rec. 1995 IEEE Int. Conf. Communications, pp. 3–8.
[15] W. D. Doyle, “Magnetization reversal in films with biaxial anisotropy,” in
1987 Proc. INTERMAG Conf., pp. 2.2-1–2.2-6.
[16] G. W. Juette and L. E. Zeffanella, “Radio noise currents n short sections on
bundle conductors (Presented Conference Paper style),” presented at the
IEEE Summer power Meeting, Dallas, TX, Jun. 22–27, 1990, Paper 90
SM 690-0 PWRS.
[17] J. G. Kreifeldt, “An analysis of surface-detected EMG as an amplitude-
modulated noise,” presented at the 1989 Int. Conf. Medicine and
Biological Engineering, Chicago, IL.
[18] J. Williams, “Narrow-band analyzer (Thesis or Dissertation style),” Ph.D.
dissertation, Dept. Elect. Eng., Harvard Univ., Cambridge, MA, 1993.
[19] N. Kawasaki, “Parametric study of thermal and chemical nonequilibrium
nozzle flow,” M.S. thesis, Dept. Electron. Eng., Osaka Univ., Osaka, Japan,
1993.
[20] J. P. Wilkinson, “Nonlinear resonant circuit devices (Patent style),” U.S.
Patent 3 624 12, July 16, 1990.
[21] IEEE Criteria for Class IE Electric Systems (Standards style), IEEE
Standard 308, 1969.
[22] Letter Symbols for Quantities, ANSI Standard Y10.5-1968.
[23] R. E. Haskell and C. T. Case, “Transient signal propagation in lossless
isotropic plasmas (Report style),” USAF Cambridge Res. Lab., Cambridge,
MA Rep. ARCRL-66-234 (II), 1994, vol. 2.
[24] E. E. Reber, R. L. Michell, and C. J. Carter, “Oxygen absorption in the
Earth’s atmosphere,” Aerospace Corp., Los Angeles, CA, Tech. Rep. TR-
0200 (420-46)-3, Nov. 1988.
[25] (Handbook style) Transmission Systems for Communications, 3rd ed.,
Western Electric Co., Winston-Salem, NC, 1985, pp. 44–60.
[26] Motorola Semiconductor Data Manual, Motorola Semiconductor
Products Inc., Phoenix, AZ, 1989.
[27] (Basic Book/Monograph Online Sources) J. K. Author. (year, month, day).
Title (edition) [Type of medium]. Volume (issue). Available: http://www.
(URL)
[28] J. Jones. (1991, May 10). Networks (2nd ed.) [Online]. Available:
http://www.atm.com
[29] (Journal Online Sources style) K. Author. (year, month). Title. Journal
[Type of medium]. Volume(issue), paging if given. Available:
http://www.(URL)
[30] R. J. Vidmar. (1992, August). On the use of atmospheric plasmas as
electromagnetic reflectors. IEEE Trans. Plasma Sci. [Online]. 21(3). pp.
876–880. Available: http://www.halcyon.com/pub/journals/21ps03-
vidmar