Hybrid of Wavelet Feature Extraction and LVQ Neural Network to

Recognize Patchouli Variety using Leaf Images

Candra Dewi

Department of Informatics, Brawijaya University, Veteran Street, Malang, Indonesia

Institute of Essential Oil, Brawijaya University, Malang, Indonesia

Keywords: Patchouli Variety, Leaf Image, Wavelet Feature Extraction, LVQ Neural Network.

Abstract: Patchouli consist of some varieties that have different patchouli alcohol (PA). This variety can be recognized

by experts who dabbling with patchouli plants through observation of shape and texture of the leaf. This study

introduced a new method to identify patchouli varieties by utilizing leaf images. The wavelet feature

extraction was used to obtain leaf texture characteristics. The varieties then are identified by using Learning

Vector Quantization (LVQ) Neural Network algorithm. The results of testing on 40 leaf image data showed

the value of recognition accuracy of patchouli varieties reached 83, 33%. This result is obtained by wavelet

parameters namely doubechies level 3, doubechies coefficient 3, and LVQ parameters, namely learning rate

0.1 learning rate reduction constant 0.2. These results can be said to be quite good considering that the

patchouli leaf tested have almost similar shape and color.

1 INTRODUCTION

Patchouli (Pogostemon cablin Benth) is one of the

essential plants that belongs to the family Labiateae.

This plant was first cultivated in the Aceh region, then

spread in several provinces such as North Sumatra,

West Sumatra, and Bengkulu. Patchouli plants

produce essential oils known as patchouli oil.

There are three types of patchouli in Indonesia

that can be distinguished by morphological character,

patchouli alcohol content (PA) and oil quality, as well

as resistance to biotic and abiotic stresses. The three

types are Pogostemon cablin Benth (Aceh patchouli),

Pogostemon heyneanus Benth (Java patchouli), and

Pogostemon hortensis Backer (Soap patchouli)

(Guenther, 1952). Of the three types, Pogostemon

cablin Benth has the highest oil content and good

composition. While Pogostemon heyneanus Benth or

Javanese patchouli more resistance to pests and

diseases, bacterial wilt and nematodes (Nuryani et al.,

1997). Besides Javanese patchouli is also resistant to

a disease, called budok in Indonesian which is caused

by the fungus Synchytrium pogostemonis (Wahyuno

and Sukamto, 2010).

Based on the description above, it can be

concluded that the selection of patchouli varieties

during crop cultivation is very necessary in order to

obtain an optimal harvest. One specific characteristic

that distinguishes patchouli varieties visually is found

in the leaves. For example, the leaves in the

Lhokseumawe variety are green and have a flat,

rounded leaf tip. While the leaves of the Sidikalang

variety are purplish green and the tips of the leaves

are flat and rounded. These differences in physical

characteristics can sometimes be recognized by

experienced of experts or farmers. However, each

variety will have different characteristics if planted in

different regions, making it even more difficult to

recognize. For example, Sidikalang varieties from

Aceh will have different leaf color and texture

characteristics if planted in Kolaka, Sulawesi. This is

often unknown to farmers and only certain experts

can recognize it. To adopt a limited number of expert

capabilities, a technology is needed in the process of

identifying patchouli leaf varieties. This paper

proposed a new method for identification of patchouli

varieties using leaf imagery. Specifically, the purpose

of this study is 1) to obtain the characteristics of leaf

texture by extracting texture features 2) to calculate

the accuracy of the recognizing of patchouli varieties

using leaf images.

Several studies on the use of leaf image

processing technology for plant identification have

been carried out. Among them is the identification of

plants through leaf shapes by counting the number of

Dewi, C.

Hybrid of Wavelet Feature Extraction and LVQ Neural Network to Recognize Patchouli Variety using Leaf Images.

DOI: 10.5220/0009954800220028

In Proceedings of the 2nd International Conference of Essential Oils (ICEO 2019), pages 22-28

ISBN: 978-989-758-456-5

leaf shape patterns, PCA, and EF (Chong et al., 2015;

Laga et al., 2014; Neto et al., 2006). Furthermore,

several studies have also been carried out for leaf

classification through texture, shape, and color

features using PSO and FRVM (Lakshmi and

Vasudef, 2016); leaf identification using DBNs and

PID (Liu and Jiang-ming, 2016); Android application

for identification of plant species based on leaf

imagery (Zhao et al., 2015); plant leaf identification

based on leaf skeleton (Zang et al., 2016);

identification of plant species based on leaf texture

(Pahikkala et al., 2015); and classification of plants

based on leaf images using backpropagation ANN

method (Aakif and M. Faisal, 2015). Other related

research is the identification of plant leaves with three

extraction features, namely shape features using the

SIFT method, color features using the color moment

method, and texture features using the SFTA method.

The use of these three features resulted in an

identification accuracy of 94% (Jamil et al., 2015).

These studies provide good enough results so that the

leaf image is quite effective for the identification of

certain plants.

In contrast to previous studies where it was used

to identify different types of plants using leaf images,

this study distinguishes plants with the same type

namely patchouli, but having different varieties. The

level of difficulty in this study lies in the

characteristics of the leaves are almost the same, so

we need an appropriate feature extraction method.

Based on field observations and discussions with

experts, it is known that almost all young patchouli

leaves have a green color and are getting red as the

plant ages. These color characteristics cannot be used

to distinguish between one variety to another. Besides

the shape of the leaves, another characteristic that can

be used to distinguish patchouli varieties is the texture

of the leaves where several varieties have slightly

different textures. To get information about leaf

texture that is almost similar requires a specific

method so that the slightest difference can be known

in detail. Of the several methods available, extraction

of texture features using wavelet texture analyzers is

one suitable alternative for patchouli leaf problems.

Wavelet ability has been demonstrated in several

studies such as in the research of Abdolmaleki et al

(2017) which extracted spectral features on

hyperspectral images and produced good

recommendations for the detection of copper

deposits. Research conducted by Bakhshipour et al.,

2017 also shows that feature extraction with wavelets

can increase the effectiveness of the weed detection

process in beet plants. Other research also shows that

the use of wavelets in feature selection can improve

performance in the recognition process (Singh et al.,

2016; Murguia et al., 2013; Imtiaz and Fattah, 2013).

In contrast to previous studies, this study used

daubechies wavelet in the transformation process.

Daubechies wavelet uses overlapping windows, so

the spectrum of high frequency coefficient represents

all high frequency changes. A daubechies level and

coefficient were also tested to get the best texture

features that can distinguish between leaf

characteristics.

The best features of each leaf image obtained from

the feature extraction process are then used as input

to the variety recognition process. This study uses the

Learning Vector Quantization (LVQ) algorithm

which is one of the Neural Network based

classification algorithms as the recognition method.

The use of the LVQ method has been done in previous

studies, namely to identify the quality of patchouli

using leaf images (Dewi et al., 2016), identification

of diseases of soybean leaves (Dewi et al., 2016;

Dewi, 2017), identification of diseases on orange

leaves (Dewi and Basuki, 2016). Research conducted

by Desylvia (2013), discusses the comparison of

SOM and LVQ in the identification of facial images

with wavelets as feature extraction. This study

concludes that the LVQ method is better than the

SOM method, with accuracy for SOM is 97.894%

and accracy for LVQ is 100% Desylva, 2013).

Furthermore, research conducted by Nurkhozin

(2011) classifies diabetes mellitus by using the LVQ

and Backpropagation method, wherein it is known

that LVQ provides a higher accuracy than

Backpropagation. The study gave 82.56% results for

LVQ and 73.25% for Backpropagation for

classification using learning rate = 0.5, number of

iteration = 100, training data were 345 and test data

were 86 patients. The above reference shows that the

use of LVQ in the identification process provides

quite optimal results.

2 DATA AND METHOD

This section gives the explanation of data and general

steps of recognizing the patchouli varieties.

2.1 Identification of Patchouli Leave

Characteristic

Patchouli is one of the plants that produce essential

oils and belongs to the Labiatea family. One of the

characteristics that can be used to identify patchouli

varieties is by observing leaf morphology (Haryudin

and Suhesti, 2014). In general, the shape of patchouli

Hybrid of Wavelet Feature Extraction and LVQ Neural Network to Recognize Patchouli Variety using Leaf Images

leaves is round and oval, with serrated leaf edges. The

shape of the tip of the leaf is pointed and leaf base is

generally blunt. Repetition of leaves almost all

pinnate accessions. The shape of the surface of old

leaves on the top of leaves is smooth wavy while the

lower surface of leaves is smooth or flat. The surface

character of the old leaf at the top is rough bumpy.

However, according to experts there are specific

characteristics on the leaves that distinguish patchouli

varieties. An example is the difference between the

Aceh patchouli and the Javanese patchouli. On Aceh

patchouli the surface of the leaves is smooth, jagged

blunt, the tip of the leaf is pointed. While the Javanese

patchouli leaves the surface of the leaves rough, the

edges of the leaves are jagged and tapered leaves.

Aceh Patchouli is more cultivated because it has

higher oil content and oil quality. This paper uses

wavelet feature extraction to obtain this texture

charactestics.

2.2

Data

The data used were taken in several regions namely

Kesamben, Brawijaya University (UB) and

Trenggalek. Data taken is image of Diploid patchouli

leaves (Kesamben and UB), Patchoulina and

Sidikalang (Trenggalek), Tetraploid (UB). Overall

data of 60 data with each variety of 10 to 20 data. An

example of patchouli leaf image is shown in Figure 1.

(a) (b) (c) (d)

Figure 1: Example images of patchouli leaf: diploid (a),

patchoulina (b), sidikalang (c), tetraploid (d).

Leaf image is taken indoors using the iPhone 4S

camera with specifications of 8 MP, f / 2.4, 35mm,

autofocus, LED flash. Leaves to be taken are placed

on a white pedestal in an upright position a distance

of 20-25 cm from the camera.

2.3

General Step of Process

General flow for the recognition of patchouli varieties

is shown in Figure 2.

The input data in the form of patchouli leaf images

as training data and test data. Furthermore, the leaf

image is processed to improve the quality by resizing

the images and converting into the gray level color

model. The resize process is carried out on the image

to equalize the pixel size of the image, which is

400x500 pixels. After that the texture extraction

process is carried out from the gray level image using

the wavelet texture analysis method.

The extracted features are Energy1 (L1) and

Energi2 (L2) then used as input to the learning

process (training) and testing process (testing).

Before testing the system, the learning process is

carried out using training data to find out the best

parameters of the LVQ algorithm, so that the best

performance is obtained. This is indicated by the

convergence of training results on the parameter

values that are learned. At the learning stage the

training data sample is used in each class as the initial

weight of LVQ. The results of the study are the

optimal final weight which is then used as the LVQ

weight in the testing process. The last stage is the

testing process on the test data using the final weights

of the results of the learning process.

After that, the accuracy calculation stage is carried

out with the aim to find out the level of accuracy of

the LVQ on identification of patchouli leaves

varieties. The results of the testing process are the

identification of the varieties that exist in the test data

image and the level of accuracy of the LVQ method.

Figure 2: General steps of recognizing process.

2.3.1 Wavelet Feature Extraction

Wavelet texture analysis is done after the matrix is

transformed using wavelet transforms. In this study

wavelet daubechies are used for transformation. The

Daubechies wavelet family is written in dbN where N

ICEO 2019 - 2nd International Conference of Essential Oil Indonesia

is the order of wavelet with a filter length of 2N and

the number of vanishing moments and db is the short

name of wavelet (Gupta, 2015). Daubechies wavelet

transforms perform calculations using leveling

decomposition and subtraction through scalar

products with proportional signals. This wavelet type

has a balance of frequency response but has a form of

nonlinear response. Daubechies wavelet uses

overlapping windows, so the high frequency

coefficient of the spectrum represents all high

frequency changes. DbN handles problems at the

edge of the data when overlapping windows by

treating the data set as if the data were periodic. The

initial sequence of data repeats by following the end

of the sequence and the end of the data is taken for the

prefix (Ian, 2001).

The basic idea of Wavelet Texture Analysis is to

extract textural features from the detail coefficient of

wavelet (sub-band) or sub picture of each

magnification. The approximate value of the sub-

band coefficient is usually represented by lighting or

image illumination variations. Thus, the framework

of the majority of wavelet texture analysis features is

extracted from high sub-bands (HH) frequencies.

By using assumption that the energy distribution

in the frequency domain can recognize textures, the

computing of energy of the wavelet sub band will

result the texture features of the image. The

calculation of texture features obtained from the

normalization of first energy (L1) or second energy

(L2) can be done using equation 1 and equation 2.













∑∑























(1)

where is 1,,,













∑∑



























/

(2)

where is 1,,,

L1 and L2 are the two energy values of the texture

projection in the subspace with Wavelet coefficient w

(i, j) at level l for sub band k, J refers to the maximum

decomposition level with horizontal wavelet

transform (h), vertical (v) and diagonal (d) on high

frequency sub band. M x N is a measure of the

coefficient of the matrix. Because the matrix is of the

same size, the value of M is equal to the value of N.

The extracted features are Energy1 (L1) and

Energi2 (L2) then used as input to the learning

process (training) and testing process (testing).

2.3.2 Leaning Vector Quantization

The Learning Vector Quantization (LVQ) is one of

the algorithms in Neural Network that perform

supervised learning against several competitive

layers. Automatically, the competitive layer learns to

group the given input vectors. Suppose there are N

data, with M input variables, and K class dividing the

data, then the steps from LVQ can be described as

follows:

1. Define:

a. Initial weights (W

) from input variable j that

falls into class k, where k is class 1 to K and j

is variables 1 through M.

b. Maximum epoch (maxEpoch) or maximum

iteration.

c. Learning rate value (α).

d. Reduction value of learning rate (decα).

e. The minimum value of learning rate that is

tolerated(minα).

2. Enter:

a. Data input (X

), where i is data 1 through N

and j is atribute 1 through M.

b. Class or target or expected output value (T

)

of each input data (X

), where i is data

through N.

3. Set the initial conditions of epoch = 0.

a. Data input (neuron input): Xij, dimana i

adalah data 1 sampai N dan j adalah variabel

1 sampai M.

b. Kelas atau Target atau nilai ouput harapan

dari masing-masing data input (Xij): Ti,

dimana i adalah data 1 sampai N.

4. Repeat the following steps if epoch epoch <=

maksEpoch dan alfa >= minAlfa:

a. Epoch value plus 1

b. Repeat the following steps from i = 1 to N:

i. Determine the value of J

obtained

from the calculation of distance

between X

and

(Jk = || Xij-

Wkj||), where k is class 1 to K.

ii. Determine the output value (Ci),

which contains the class of initial

weights (W

) which has the smallest

or minimum J (Ci = minimum J

iii. Update the initial weight (W

) with

the following provisions:

 If Ti =Ci, then

)(

kjijkjkj

WXalfaWW 



(6)

 If Ti <> Ci, then

)(

kjijkjkj

WXalfaWW 



(7)

Hybrid of Wavelet Feature Extraction and LVQ Neural Network to Recognize Patchouli Variety using Leaf Images

c. Reduce the α value, by means of α = α - (α

* decα) or α = α -decα

After the training process, the final weights

) are obtained and the values are used to

perform the testing.

3 RESULT AND DISCUSSION

The training phase with training data is carried out to

get optimal parameters both for Wavelet parameters

and LVQ parameters. The Wavelet parameters tested

were doubechies coefficient (db coefficient) and

doubechies level (db level), while the LVQ

parameters tested were learning rate and learning rate

reduction. The results of testing these parameters are

shown in Table 1, Table 2 and Table 3. The training

processes was carried out with an LVQ iteration of

1000. Furthermore, the LVQ weight obtained in the

training process was used to conduct the test on the

training data and test data.

3.1 The Experiment of Doubechies Level

and Coefficient

The db coefficient and db level tests were performed

at a learning rate of 0.1 and a reduction in learning

rate of 0.1. The db level tested were 1, 2 and 3, while

the db coefficient tested ranged from 1 to 10.

The test results at Table 1 show the best db level

was 3 and the best db coefficient was 2, 3, 4 and 10.

Both training and test data produces the same

accuracy value of 83.33%.

Table 1: The result of db level and db coefficient test.

Level

Coefficient

Accuracy (%)

Train data Test data

1 1 70,8 58,3

2 75 66,7

3 75 66,7

4 75 66,7

5 79,2 66,7

6 79,2 66,7

7 75 66,7

8 75 66,7

9 70,8 66,7

10 75 66,7

1 79,2 75

2 83,3 66,7

3 83,3 66,7

4 83,3 66,7

5 79,2 75

Level

Coefficient

Accuracy (%)

Train data Test data

6 79,2 66,7

7 79,2 66,7

8 75 66,7

9 75 66,7

10 79,2 66,7

3 1 83,33 75

2 83,33 83,33

3 83,33 83,33

4 83,33 83,33

5 83,33 75

6 79,2 75

7 79,2 75

8 79,2 75

9 79,2 75

10 83,33 83,33

3.2 The Experiment of Learning Rate

The best parameter values of Wavelet obtained from

the test are then used as a reference in testing the LVQ

parameters. The learning rate test was performed at

db level 3, db coefficient 3 and the learning rate

reduction is 0.1. This test is carried out at learning rate

ranging from 0.1 to 0.9. The result of the learning rate

test is shown in Table 2.

Table 2: The result of learning rate test.

Learning

rate

Accuracy (%)

Train data Test data

0,1 83,33 83,33

0,2 83,33 75

0,3 83,33 75

0,4 83,33 75

0,5 83,33 75

0,6 83,33 75

0,7 87,5 75

0,8 45,8 41,7

0,9 41,7 33,3

The test results show the best accuracy for training

data is 87.5% at learning rate 0.7 and for testing data

was 83.33% for testing data at leaning rate 0.1.

However, the most optimal accuracy for both training

data and test data that was equal to 83.33% at leaning

rate 0.1.

3.2 The Experiment of Learning Rate

Reduction

The learning rate reduction test uses level db 3,

coefficient db 3 and the learning rate value 0.1. The

result of the learning reduction test is shown in Table

3. The test results show that the best learning rate

ICEO 2019 - 2nd International Conference of Essential Oil Indonesia

reduction is 0.2, and 0.4 with an accuracy value of test

data is 83.33% (Table 3).

Table 3: The result of learning rate reduction test.

Learning rate

reduction

Accuracy (%)

Train data Test data

0,1 83,33 83,33

0,2 87,5 83,33

0,3 87,5 75

0,4 87,5 83,33

0,5 83,33 83,33

0,6 79,2 83,33

0,7 75 83,3

0,8 75 83,3

0,9 70,8 83,3

4 CONCLUSIONS

This study carried out the identification of patchouli

plant varieties using the image of patchouli leaves.

This process combines the ability of the wavelet

method to extract texture features and LVQ for the

classification of patchouli varieties. The process of

identifying patchouli varieties begins with the

training to get the optimum wavelet parameters (db

level and db coefficient) and LVQ parameters

(constant of learning rate and learning rate reduction)

to find out the optimal method performance. Test

results at db level 3, db coefficient 2, 3 and 4, learning

rate 0.1 and the reduction of leaning rates 0.2 and 0.4

obtained the highest accuracy is 83.33%. The results

obtained are quite good, but further research needs to

be done especially by increasing the amount of data

and adding patchouli varieties.

ACKNOWLEDGEMENTS

We would like to thank to Faculty of Computer

Science, University of Brawijaya for the funding of

this research.

REFERENCES

Aakif, A., Faisal Khan, M., 2015. Automatic Classification

of Plants Based on Their Leaves. Biosystems

Engineering. 139, 66–75.

Abdolmaleki, M., Tabaei, M., Fathianpour, N., Gorte, B.

G.H., 2017. Selecting Optimum Base Wavelet For

Extracting Spectral Alteration Features Associated

With Porphyry Copper Mineralization Using

Hyperspectral Image. International Journal of Applied

Earth Observation and Geoinformation, 58, 134-144.

Bakhshipour, B., Jafari, A., Nassiri, S. M., Zare, D., 2017.

Weed Segmentation using Texture Features Extracted

from Wavelet Sub-Images. Biosystems Engineering,

157, 1-12.

Zhao, C., Chan, S.S.F., Cham, W.K., Chu, L.M., 2015.

Plant Identification using Leaf Shapes—A Pattern

Counting Approach. Pattern Recognition. 48, 10,

3203–3215

Dewi, C, Krisnanti, G.W., Cholissodin, I., Basuki, A., 2016.

Identifying Quality of Patchouli Leaves through Its

Leave Image Using Learning Vector Quantization. The

Annual Basic Science International Conference,

March 2016, Malang, Indonesia.

Dewi, C., Umam, M. S., Cholissodin, I., 2016.

Identification of Disease on Leaf Soybean Image Using

Learning Vector Quantization. International Congress

on Engineering and Information, May 2016, Osaka,

Japan.

Imtiaz, H., Fattah, S. A., 2013. A Wavelet -Based

Dominant Feature Extraction Algorithm for Palm-Print

Recognition. Digital Signal Processing, 23(1), 244-

258.

Jamil, N., Aslina, N., Hussin, C., Awang, K., 2015.

Automatic Plant Identification: Is Shape the Key

Feature?. Procedia Computer Science, 76, 2015, 436-

442.

Lakshmi, B.V., Mohan, F., 2016. Kernel-Based PSO and

FRVM: An Automatic Plant Leaf Type Detection using

Texture, Shape, and Color Features. Computers and

Electronics in Agriculture, 125, 99–112.

Liu, N., Kan, J-M., 2016. Improved Deep Belief Networks

and Multi-Feature Fusion for Leaf Identification.

Neurocomputing, 216, 460–467.

Laga, H., Kurtek, S., Srivastava, A., Miklavcic, S.J., 2014.

Landmark-Free Statistical Analysis of the Shape of

Plant Leaves. Journal of Theoretical Biology, 363, 41–

52.

Murguía, J.S., Vergara, A., Vargas-Olmos, C., Wong, T. J.,

Fonollosa, J., Huerta, R., 2013. Two-dimensional

Wavelet Transform Feature Extraction for Porous

Silicon Chemical Sensors. Analytica Chimica

Acta, 785, 1-15.

Neto, J. C., Meyer, G.E., Jones, D.D., Samal, A.K., 2006.

Plant Species Identification using Elliptic Fourier Leaf

Shape Analysis. Computers and Electronics in

Agriculture. 50(2), 121–134.

Pahikkala, T., Kari, K., Mattila, H., Lepisto,

A., Teuhola, J., Nevalainen, O.S., Tyystjärvi, E., 2015.

Classification of Plant Species from Images of

Overlapping Leaves. Computers and Electronics in

Agriculture, 118, 186–192.

Singh, A., Dutta, M. K., Sarathi, M.P., Uher, V., Burget, R.,

2016. Image processing Based Automatic Diagnosis of

Glaucoma using Wavelet Features of Segmented Optic

Disc from Fundus Image. Computer Methods and

Programs in Biomedicine, 124, 108-120.

Zhang, L., Weckler, P., Wang, N., Xiao, D., Chai, X.,

2016. Individual Leaf Identification from Horticultural

Hybrid of Wavelet Feature Extraction and LVQ Neural Network to Recognize Patchouli Variety using Leaf Images

Crop Images Based on the Leaf Skeleton. Computers

and Electronics in Agriculture, 127, 184–196.

Zhao, Z.-Q., Ma, L.H., Cheung, Y.M, Wu, X., Tang, Y.,

Chen, C.L.P., 2015. ApLeaf: An Efficient Android-

Based Plant Leaf Identification System.

Neurocomputing, 151(3), 1112–1119.

Haryudin, W., Suhesti, S., 2014. Karakteristik Morfologi,

Produksi dan Mutu 15 Aksesi Nilam (Morphological

characteristics, production and quality of 15 patchouli

accessions). Bul. Littro, 25(1).

ICEO 2019 - 2nd International Conference of Essential Oil Indonesia