A Low Cost Electronic Nose with a GMM-UBM Approach for Wood

Species Veriﬁcation

Naren Mantilla-Ramirez

, Homero Ortega-Boada

1 a

, Milton Paja-Sarria

2 b

and Alexander Sep

ulveda-Sep

ulveda

1 c

Escuela de Ingenier

ıas El

ectrica, Electr

onica y de Telecomuniciones (E3T), Universidad Industrial de Santander,

Bucaramanga, Colombia

Facultad de Ingenier

ıas, Universidad Santiago de Cali, Cali, Colombia

Keywords:

Wood Identiﬁcation, Timber Identiﬁcation, Chemical Sensor Arrays, GMM-UBM, Gaussian Mixture Models,

Universal Background Model.

Abstract:

Deforestation endangers some vulnerable wood species. Although there are effective timber species identiﬁ-

cation methods, they are typically expensive and time-consuming, they must be carried out by experts and they

are not applicable to places far from main cities. In contrast, we propose to use electronic noses to identify

timber species, e.g. during their transportation process, from the volatile compounds that timbers emanate. In

the present work, it is proposed a method for timber species detection from their aromas. The measurements of

the volatile compounds are made by an array of 16 chemical sensors, whose curves are the inputs to a pattern

recognition system. Detection is performed by using Gaussian mixture modeling with Universal Background

Model. In contrast to previous works, in this work, we apply a new approach to the problem of timer species

detection; furthermore, the sample collection conditions are closer to those found in real situations; and, the

number of samples used is larger and more varied. We found an EER (equal error rate) of 24.18% for cedar

veriﬁcation and an EER of 33.62% for 4-timber species veriﬁcation.

1 INTRODUCTION

Deforestation occurs all around the globe, but espe-

cially in tropical countries like Colombia, where il-

legal logging is one of its main causes. Due to this

process, and due to the increasing timber demand,

particular tree species that can be found in wild areas

are threatened. Despite efforts to protect the countrys

natural resources, Colombian entities still struggle to

combat illegal logging. In fact, deforestation rates in

Colombia remain notably high, especially in recent

years. In this regard, a common procedure carried by

the police, consists on stopping a truck transporting

timber in order to ask the driver for a letter of safe

passage. Then, the timber cargo is veriﬁed; however,

appropriate monitoring instruments are required in or-

der to detect timber from vulnerable and prohibited

tree species.

Among different wood identiﬁcation strategies,

https://orcid.org/0000-0003-0343-8625

https://orcid.org/0000-0003-4288-1742

https://orcid.org/0000-0002-9643-5193

looking at macroscopic features such as color, texture

and odour, stands out because they can be used to es-

tablish whether a wood is correctly labeled (Wheeler

and Baas, 1998), which is highly convenient when an-

alyzing large volumes of timber. Using these macro-

scopic features, by part of trained personnel, is the

most common way of timber identiﬁcation in Colom-

bia, but it is done empirically and subjectively.

There are also precise methods based on taxo-

nomic and genetic analyzes, in which wood species

samples are compared at using genetic sequencing

techniques (Hanssen et al., 2011; Yu et al., 2016).

Although the reliability of these tests is almost 100%,

they are expensive, delayed and they must be carried

out by experts. Other used techniques involve differ-

ent spectroscopic (Rana et al., 2008; Cabral et al.,

2012) and imaging (Dickson et al., 2017) analyzes,

but they still require support from experts and a con-

siderable amount of time. All previously mentioned

approaches are effective techniques, but they do not

still meet the requirements to be applied in suburban

and rural regions away from major cities (Kalaw and

Sevilla, 2018).

Mantilla-Ramirez, N., Ortega-Boada, H., Paja-Sarria, M. and Sepúlveda-Sepúlveda, A.

A Low Cost Electronic Nose with a GMM-UBM Approach for Wood Species Veriﬁcation.

DOI: 10.5220/0008978003330341

In Proceedings of the 9th International Conference on Pattern Recognition Applications and Methods (ICPRAM 2020), pages 333-341

ISBN: 978-989-758-397-1; ISSN: 2184-4313

333

As an alternative approach, it has been pro-

posed to analyze volatile compounds emitted by wood

species by using strategies such as gas chromatogra-

phy (Fedele et al., 2007; M

uller et al., 2006; Rinne

et al., 2002), but this approach is expensive as well.

On the other hand, a less expensive and practical

option is to use electronic noses (e-nose) (Kalaw and

Sevilla, 2018; Wilson, 2012). E-nose systems have

been used in a growing number of applications, for

example in food industry (Shi et al., 2017), air quality

analysis (Capelli et al., 2014), explosives (Guo et al.,

2017), and narcotics detection, among others (San-

tos and Lozano, 2015). It has also been proposed

these devices to be used for timber species identi-

ﬁcation from the volatile compounds that they em-

anate (Cordeiro et al., 2016; Kalaw and Sevilla, 2018;

Wilson, 2012).

In particular, the authors in (Garneau et al., 2004)

analyzed three different species of Pinaceae family

by using electronic noses; where, Principal Compo-

nent Analysis (PCA) was performed, showing ob-

servable differences between the species. The data

set was composed by 30 smell-prints acquired from

wood samples of the three conifer species. Later in

(Wilson et al., 2005), the authors claim to have used

neural networks as a mechanism to identify wood

species, where identiﬁcation rates between 94% and

99% were obtained. Regarding the dataset, two sam-

ples per tree, belonging to 13 − 30 trees of 12 dif-

ferent wood species, were taken. Another study car-

ried out in Brazil (Cordeiro et al., 2016), classiﬁca-

tion with electronic noses for two pairs of woods was

performed: (a) mahogany vs Spanish-cedar; and, (b)

Brazilian walnut vs black-cinnamon. The calculated

relative responses from the electronic nose were used

as input data for principal component analyses (PCA),

showing satisfactory results. However, details about

data set composition is not provided. Finally, the au-

thors in (Kalaw and Sevilla, 2018) used gas sensors

measurements to analyze the separability of ﬁve tim-

ber species in the Philippines. The authors reported

separable clusters at ﬁrst glance when features ob-

tained by principal component analysis (PCA) were

used.

Notwithstanding the good results reported in pre-

viously mentioned studies, the experiments were per-

formed by using a reduced amount of samples and

in controlled conditions, far from practical situations.

This work aims to advance in this regard and present

an approach which can work in a less controlled en-

vironment, moving the experiment away from ideal

conditions. In Colombia, the extraction of wood

occurs mainly in remote regions of difﬁcult access,

which hinders the constant presence of experts, the

transportation of wood samples and the installation of

specialized equipment. Therefore, a solution adapt-

ing to practical conditions is required. In particular,

in present work we use timber samples after trans-

portation procedures, instead of wood material. The

reason to do this is because the aromas are very fresh,

strong, and without major interference when working

with wood material, quite the opposite occurs in tim-

ber identiﬁcation.

In addition to this, previous research in this area

report results in identiﬁcation tasks instead of ver-

iﬁcation. The main difference between identiﬁca-

tion and veriﬁcation tasks, is that the former case

is a N-class classiﬁcation problem, whilst later one

is a binary classiﬁcation problem. This is a widely

discussed topic, specially in speaker recognition re-

search (Doddington et al., 2000). Veriﬁcation authen-

ticates an individual timber sample by comparing it

with one speciﬁc biometric reference stored in the

database, while identiﬁcation compares it with all the

bio-metrics stored in the database. As far as we know,

this is the ﬁrst work in which timber species detection

procedure is performed by the smell, from a biometric

veriﬁcation point of view.

This paper proposes just that. Here, we propose

a method for timber species veriﬁcation from their

aromas. Timber samples were collected from wood

deposits in the “Gran Santander” region in Colom-

bia, making data collection closer to practical situa-

tions. The measurements of the volatile compounds

were made by using an array of 16 chemical sensors,

whose curves are the inputs to a pattern recognition

based system. Data pre-processing is performed using

Principal Component Analysis (PCA) (Jolliffe, 2011)

and, the detection is carried out by using Gaussian

mixture modeling with Universal Background Model

(Reynolds et al., 2000).

2 METHOD

2.1 Electronic Nose System

Figure 1 shows the building blocks for an e-nose sys-

tem (Ruiz Jim

enez, 2018). The chemical phase cor-

responds to the sensing of volatile compounds by the

gas sensors array. In electronic phase, electrical sig-

nals are acquired and conditioned in order to obtain a

temporal matrix representation of the sensed sample.

Finally, this data is processed by pattern recognition

algorithms in order to detect the tree species of timber

material.

There are different types of gas sensors, classiﬁed

according to their size, sensitivity, application, and

ICPRAM 2020 - 9th International Conference on Pattern Recognition Applications and Methods

334

Che

Volatile

Conditioning

Sensors

Array

Signal

Conditioning

Pattern

Recognition

and analysis

Chemical Phase Electronic Phase Software

Volatile

compounds

Identified

odor

Figure 1: Typical E-NOSE general scheme. Adapted from

(Ruiz Jim

enez, 2018).

technology. Sensors based on metal oxide semicon-

ductor ﬁlms are the most common. They are com-

posed by n-type metal oxide crystals, such as tin(IV )

oxide (SnO

, aka stannic oxide). These sensors sensi-

tivity varies with temperature, so they usually have

a heating element controlled by an electric current

for the sake of keeping the temperature in a constant

value. Also, before being used for the ﬁrst time, the

sensors must go through a preheating stage (Figaro,

2005). Despite these inconveniences, these sensors

are preferred because they have stable characteristics

over time and do not require frequent maintenance

processes (Ghasemi-Varnamkhasti et al., 2019).

Regarding the many electronic noses available in

the literature, there are commercial electronic noses,

whose sensors consist of nonconducting organic poly-

mers. Several types have been used such as the Cyra-

nose 320 (Garneau et al., 2004) and the Aromascan

A32S (Wilson et al., 2005). However, this type of

devices do not meet the research needs, because it

has ﬁxed, unmodiﬁable and non-customized sensor

arrays.

On the other hand, customized e-noses have been

used, for example: in (Kalaw and Sevilla, 2018), an 8-

chemical sensors array with resistive principles based

on carbon nanotubes was used; and in (Cordeiro et al.,

2016), a 4-conductive polymer sensors array with re-

sistive principle was utilized. However, the manufac-

ture of sensors is out of the scope of present work.

As an alternative to the aforementioned, the elec-

tronic nose used in herein corresponds to a proto-

type developed in Universidad Industrial de San-

tander (Ruiz Jim

enez, 2018), as shown in Fig. 2. It

was developed under the DIY (do it yourself, do it

yourself) culture, which allows to conduct research

at different scales and at a low cost. It is com-

posed by a 4 × 4 matrix array sensor. The signal

conditioning and acquisition process is carried out

by using an Intel Galileo Generation 1 acquisition

board (Ruiz Jim

enez, 2018). This prototype has semi-

conductor metal oxide sensors, which vary their elec-

trical resistance due to the chemical reaction that oc-

curs when gases make contact with the sensors. These

sensors belong to the manufacturing houses Figaro

Engineering and Hanwei Electronics, which are char-

acterized by their ability to detect low gas concentra-

tions and by their low cost. Table 1 lists the sensors

used in the prototype.

Figure 2: E-NOSE prototype developed by UIS

in (Ruiz Jim

enez, 2018).

2.2 Data Collection Procedure

In this work, a total of 309 samples (woodblocks) of

different wood species were taken from wood tim-

ber stocks of different cities in the Gran Santander

region in Colombia (Bucaramanga, Lebrija, Socorro,

San Gil, Pamplona, and Ccuta).

Before taking samples and perform all the mea-

surement experiment, it was necessary to develop two

previous tasks: e-nose preparation and sample prepa-

ration. First, the e-nose is turned on for one hour

so that the sensors reach their steady-state operation

in the corresponding environment. Later, each sam-

ple (woodblock) is prepared by brushing it 20 times

with a wooden brush, and the resulting material is dis-

carded to eliminate possible contamination by contact

with another sample, or interference with other ele-

ments. Then perform the following procedure as de-

scribed below for each wood sample:

• Brush the sample 20 more times and take approx-

imately 1 cm

of the resulting wood chip.

• Sense (sniff) the sample with the e-nose. As a

result, 16 response curves, corresponding to the

conductance variations of each sensor in the ma-

trix array, are obtained. This group of curves is

known as the smell-print of the wood sample.

• Let the sensors rest during 5 minutes, allowing

the entry of airﬂow generated by a fan, in order

to avoid previous trials interfere with the current

trial.

Each response curve was taken at a sampling pe-

riod of 270 ms, following three steps. First, the sen-

sors react to air for 100 samples; then, the correspond-

ing wood chips are placed inside sensors chamber

A Low Cost Electronic Nose with a GMM-UBM Approach for Wood Species Veriﬁcation

335

Table 1: Sensors in the E-NOSE prototype.

SENSOR BRAND REF SENSOR BRAND REF

1 HANWEI MQ-2 9 FIGARO TGS-832

2 HANWEI MQ-3 10 HANWEI MQ-6

3 HANWEI MQ-4 11 FIGARO TGS-823

4 HANWEI MQ-6 12 FIGARO TGS-816

5 HANWEI MQ-7 13 FIGARO TGS-822

6 HANWEI MQ-8 14 FIGARO TGS-813

7 HANWEI MQ-135 15 FIGARO TGS-826

8 HANWEI MQ-9 16 HANWEI MQ-3

during 300 samples; ﬁnally, the wood chips are re-

moved and the sensors are exposed to the air again

with residual air for additional 100 samples. In ﬁgure

3, it is shown an example curve response of one of the

sensors.

Phase 1 Phase 2 Phase 3

Conductance

Samples at 3.7 Hz

100 400 500

Figure 3: The different stages of the response are shown.

Phase 1, ﬁrst 100 frames, where the base reading is done.

Phase 2, frames from 101 to 400, where sensors capture the

smell of the wood chip sample. Phase 3, frames from 401 to

500, where sensors begin to return to their reference state.

To reduce the electronic noise effect in the acqui-

sition system, a ﬁltering process is carried out using

an ﬁfth-order median ﬁlter for each sensor response

curve. It allows to reduce outliers inﬂuence on the

measurement process.

2.3 Feature Extraction

In this paper, two strategies were used for the feature

extraction stage. In the ﬁrst, heuristic parameters de-

scribing the behavior of each curve were calculated

as reported in recent works. Second case scenario,

the recognition process is performed on the raw data.

2.3.1 Heuristic Parameters

Different features can be estimated from the response

curve of the conductance value of each sensor. In pre-

vious works, it is reported the use of the maximum

and minimum values, and, area under the curve as

features (Yan et al., 2015). Another way is to use

strategies that involve a transient analysis of the sen-

sor’s response (Rodriguez-Lujan et al., 2014; Rana

et al., 2008). Finally there is a third way, where a pre-

deﬁned model is adjusted using available data (data

driven model) (Carmel et al., 2003). In the present

work the following features were estimated:

• G

, initial conductance value, mean of the ﬁrst

100 samples of the total response.

• G

, ﬁnal conductance value, mean of the last 50

samples of the phase 2 of the total response.

• G

max

, maximum conductance value.

• G

min

, minimum conductance value.

• B, gain coefﬁcient; and A, pole location, corre-

sponding to an adjusted ﬁrst-order auto-regressive

model:

H(z) =

−1

1 + Az

−1

. (1)

In previously reported research works, principal

component analysis (PCA) is typically used to decor-

relate variables and reduce dimensionality and then

avoid over-ﬁtting (Akbar et al., 2016; Cordeiro et al.,

2016). PCA analysis allows to reduce the dimension-

ality by applying a linear transformation that maps

the data into a new space, where the new variables

are uncorrelated. The new dimensions are sorted by

the amount of variance they describe, from highest

to lowest, concentrating the largest amount of vari-

ance in the ﬁrst components (Jolliffe, 2011). In this

way, by taking a few main components it is possible

to represent most the variability present in the original

data. Herein, after pilot experiments, we uses 90% as

the level of variance to be represented by the p PCA

components.

2.3.2 Full Time Series

For these experiments, we consider each of the N =

16 sensors as a single characteristic, resulting in a fea-

ture vector of dimension N. Furthermore, we include

the ﬁrst s = 400 samples of each curve response. By

ICPRAM 2020 - 9th International Conference on Pattern Recognition Applications and Methods

336

using the ﬁrst s samples, it is not necessary to reduce

the dimension, because we are using almost all the in-

formation to estimate the GMM model, i.e., 400 fea-

ture vectors of length 16 per timber block.

2.4 GMM-UBM Approach for Timber

Species Detection

The aim of this work is to support authorities in their

ﬁght against illegal and selective timber species log-

ging. In this regard, an identiﬁcation approach typi-

cally uses a closed set to classify which among N pos-

sibilities is most probable. On the other hand, species

veriﬁcation may be a better approach to make the er-

ror rate drops. Veriﬁcation procedures allow deter-

mining whether a sample belongs to a class by having

enough enrolled samples from that species.

A Universal Background Model (UBM) is a con-

cept taken from biometrics that, in this case, corre-

sponds to timper species-independent model, repre-

senting the universe or expected overall evaluation

conditions. Regarding the training data for the UBM,

selected samples should reﬂect the expected alter-

native hypothesis to be encountered during recogni-

tion. The veriﬁcation task can be summarized in test-

ing whether a sample corresponds to the analyzed

class or another unknown class (alternative hypoth-

esis). In this case, the impostor hypothesis (any other

class) is modeled by the Universal Background Model

(UBM) (Reynolds et al., 2000; Doddington et al.,

2000).

In this work, two veriﬁcation tasks were per-

formed: ﬁrst, one class (cedar) vs the universal model,

thus individuals belonging the rest of species are in-

cluded in the UBM model; and second, each of four

particular classes (cedar, moncoro, pine and sapan) vs

the universal model, therefore, these 4 model are not

included in the UBM model. In table 2, there is a

description of the number of individuals per class.

The UBM models a probability density function

(PDF) that represents the properties of the reference

smell-print species population. In that sense, the

doubtful smell-print is compared with respect to the

UBM as well as a PDF model of the a particular

timber species. In such case, there are two models:

timber species model (λ

) and Reference UBM (λ

When passing the observations corresponding to the

intercepted signal X two probability values, p(X |λ

)

and p(X |λ

), are obtained, with which Likelihood

Ratio (LR) is built. However, it is common to use

the Log Likelihood Ratio (LLR),

L(X ) = log p(X |λ

) − log p(X |λ

). (2)

As the value L(X ) increases, the evidence that the

doubted smell-print correspond to the species we are

looking for becomes stronger.

For PDF modelling, the well known Gaussian

mixture model is preferred. The use of Gaussian mix-

ture models is motivated by their capability to model

arbitrary densities (Kinnunen and Li, 2010; Reynolds

and Rose, 1995). A GMM is composed of a ﬁnite

mixture of multivariate Gaussian components and the

set of parameters denoted by λ. It is characterized by

a weighted linear combination of C unimodal Gaus-

sian densities by the function:

p(o

λ ) =

∑

i=1

N (o, µ

, Σ

), (3)

where o is a D-dimensional observation or feature

vector, α

is the mixing weight (prior probability)

of the i-th Gaussian component, and N (·) is the D-

variate Gaussian density function with mean vector µ

and covariance matrix Σ

. The popular expectation-

maximization (EM) algorithm is used for maximizing

the likelihood with respect to a given data. The in-

terested reader is referred to (Bishop, 2006; Reynolds

et al., 2000; Kinnunen and Li, 2010) for more com-

plete details.

3 RESULTS

In order to measure the performance of the proposed

timber species detection system, the DET (Detection

Error Trade-off) curves and the EER (Equal Error

Rate) value are used. DET curves plots False Rejec-

tion Rate (FRR, in the Y-axis) versus False Accep-

tance Rate (FAR, in the X-axis), where the curve that

is closest to the left bottom corner of the graph cor-

responds to the system having the best performance

(Martin et al., 1997).

The veriﬁcation experiments included 309 sam-

ples from at least 18 wood species, as shown in ta-

ble 2. First, 67 cedar samples were clustered in a

class, and 180 samples from other species were used

to ﬁt the UBM with a 4-Gaussian mixture model. The

rest of the samples were used for validation: 17 cedar

samples and 45 impostors. We called cedar detec-

tion to this experiment. In the second experiment,

which we named 4-species detection, 159 wood sam-

ples were clustered into four classes (67 cedar sam-

ples, 37 moncoro samples, 21 pine samples, and 34

sapan samples). Again, samples of the other species

(86 samples) were used to ﬁt the UBM with a 4-

Gaussian mixture model. The rest of the samples

were used for validation: 17 cedar samples, 10 mon-

coro samples, 6 pine samples, 9 sapan samples and

A Low Cost Electronic Nose with a GMM-UBM Approach for Wood Species Veriﬁcation

337

Table 2: Samples per species used in this work.

Wood Species Scientiﬁc Name Number of wood blocks

Cedar Cedrela odorata 84

Moncoro Cordia gerascanthus 47

Pine Retrophyllum rospigliosii 27

Sapan Clathrotropis brunnea 43

Others

Tabebuia aurea, Zanthoxylum rhoifolium, Fraxinus

uhdei, Anacardium excelsum, Simarouba amara,

Cariniana pyriformis, Ficus spp., Quercus humboldtii,

Guarea guidonia, Coffea arabica, Alchornea

triplinervia, Corymbia citriodora, Swietenia

macrophylla, and others unknown.

108

22 impostors. For the ﬁrst, as well as for the sec-

ond experiment, the PDF of each class representing a

particular timber species is modeled by a 4 Gaussian

mixture model. The EER value is estimated by using

a cross-validation procedure of 5-sets; where 80% of

timber samples corresponds to the training set, whilst

the remaining samples (20%) were used for valida-

tion. Each veriﬁcation problem (one class vs. UBM,

and four clases vs. UBM) were analyzed from two

scopes: ﬁrst, by using traditional feature extraction

methods followed by PCA dimensionality reduction;

and second, considering the full time series of the 16

curves.

3.1 Cedar Detection

In the ﬁrst set of experiments, we apply a feature

extraction procedure, and six values were calculated

from each of the 16 curves. This process results in

a 96-dimensional feature vector for each sample, that

is, an X

309×96

matrix. In previously reported works,

principal component analysis (PCA) is typically used

to reduce the dimension and avoid over-ﬁtting (Ak-

bar et al., 2016; Cordeiro et al., 2016). In the case

of this application, ﬁve principal components were

used to represent approximately 90% of the variance

in the original data. On the other hand, the validation

samples were compared against a single model of the

known class (cedar) and the UBM. In this way, it is

veriﬁed whether the analyzed sample belongs to the

cedar class or not. The veriﬁcation with one class,

applying PCA, showed a classiﬁcation error rate of

38.49% with a standard deviation of 7.02%. Figure 4

depicts the DET curve for this experiment.

Next, we consider the signal acquired by each of

the 16 sensors as a characteristic, forming a feature

16-dimensional feature vector. Then, each of the 400

samples obtained is considered as a frame, thus our

representation of a wood block smell-print is a 16 ×

400 matrix.

The results of the cedar class veriﬁcation proce-

5 7 10 15 20 30 40 50 60

False Positive Rate (FPR) [%]

False Negative Rate (FNR) [%]

Figure 4: DET (Detection Error Trade-off) curve for one

class veriﬁcation with Principal Component Analysis.

dure vs. the UBM, using the full time series, showed

a classiﬁcation error rate of 24.18% with a standard

deviation of 7.71%. A DET (Detection Error Trade-

off) curve is shown in Figure 5.

3.2 4-Species Detection

The validation samples were compared with the four

models of the known classes (Cedar, moncoro, pine,

and sapan) and the UBM, and the LLR values are cal-

culated. In this way, it is veriﬁed whether the analyzed

sample belongs to one of the four known classes or

not. The veriﬁcation with four classes, applying PCA,

showed a classiﬁcation error rate of 47.72% with a

standard deviation of 3.79%. In Figure 6, it is shown

a DET (Detection Error Trade-off) curve.

Considering each of the 16 curves response as a

time series, we apply the veriﬁcation procedure. The

results of the veriﬁcation with four classes, using the

full time series, showed a classiﬁcation error rate of

ICPRAM 2020 - 9th International Conference on Pattern Recognition Applications and Methods

338

5 7 10 15 20 30 40 50 60

False Positive Rate (FPR) [%]

False Negative Rate (FNR) [%]

Figure 5: DET (Detection Error Trade-off) curve for one

class veriﬁcation with all data.

5 7 10 15 20 30 40 50 60

False Positive Rate (FPR) [%]

False Negative Rate (FNR) [%]

Figure 6: DET (Detection Error Trade-off) curve for four

class veriﬁcation with Principal Component Analysis.

33.62% with a standard deviation of 4.14%. A DET

curve is shown in Figure 7.

In Table 3, it is presented a summary of the Equal

Error Rates (EER) for the experiments.

Table 3: Summary of Equal Error Rate for the experiments

carried out.

Target

Detection

Analysis

Mean

EER

Standard

deviation

Cedar PCA 38.49% 7.02%

Cedar Full series 24.18% 7.71%

4-species PCA 47.72% 3.79%

4-species Full series 33.62% 4.14%

5 7 10 15 20 30 40 50 60

False Positive Rate (FPR) [%]

False Negative Rate (FNR) [%]

Figure 7: DET (Detection Error Trade-off) curve for four

classes veriﬁcation with all data.

4 DISCUSSION

The reported errors in this work are high, in con-

trast with high success rates reported in (Kalaw and

Sevilla, 2018; Cordeiro et al., 2016; Wilson, 2012).

However, it is important to take into account that

previous mentioned works use wood material, whose

aromas are fresh, strong and without major interfer-

ence. In addition, those reported works use controlled

conditions. By contrast, in this paper we use timber

samples, which are less fresh and have weaker aro-

mas. Although timber samples are more difﬁcult to

classify, their conditions are closer to real situations

than those when using wood samples.

Furthermore, reference works performed wood

species identiﬁcation tasks, which involves a limited

and closed number of species within which to clas-

sify a sample. On the other hand, we propose a ver-

iﬁcation procedure, more useful in practice, in which

a sample is compared with a reference model corre-

sponding to a species of interest. If a test sample does

not resemble the target class, it is said to belong to

another class; while in the identiﬁcation processes, a

label of the deﬁned classes must be assigned.

Wood species veriﬁcation procedures are unusual,

less those approaches based on the aroma. Within the

search for information carried out for this work, there

are no reports about timber veriﬁcation by smell. Our

proposal is to combine the use of electronic noses

and veriﬁcation techniques, such as GMM-UBM, to

quickly determine whether a timber sample belongs

to a species of interest or not, based on its aroma. It

A Low Cost Electronic Nose with a GMM-UBM Approach for Wood Species Veriﬁcation

339

is important to highlight the differences with data col-

lection methods reported in previous research works

related to the area, as well as the number of samples,

origin, species, pre-process, and previous storage of

the samples. The objective of this work is to ana-

lyze a greater dataset, with non-freshly sawn wood

samples and with non-rigorous storage conditions. It

allows establishing less distant from the real environ-

ment conditions for which the problem is sought to be

solved. With this being said, results within the scope

of this study are promising, as it shows that analyzed

signals contain important and discriminative informa-

tion for the task at hand.

Although it is used a larger number of samples

than in other reference works, the dataset is still non-

large enough. Furthermore, the approach of using a

veriﬁcation scenario seems to be appropriate when

there is interest of targeting speciﬁc species in an en-

vironment where there could be some non-identiﬁed

species which are hard to label. In addition to this, it

would be necessary to explore other feature extrac-

tion techniques, for example, by exploring spectral

information using different time-frequency represen-

tations, and multi-channel approaches such as those

used in electroencephalogram (EEG) signal analysis.

As could be notice in the results presented herein, the

PCA approach performs worst that using the time se-

ries directly. It is also better to verify whether a sus-

picious sample belongs to a single species or not.

5 CONCLUSIONS AND FUTURE

WORK

It was proposed a method for smell-based wood

species detection by using a low cost electronic nose,

which is formed by an array of 16 chemical sen-

sors. Veriﬁcation was carried out by using Gaussian

mixture modeling with Universal Background Model.

Veriﬁcation procedures are a better option in practical

scenarios than identiﬁcation procedures. As far as we

know, our work is the ﬁrst approach that made use

of smell-prints from a biometric approach for wood

species detection .

Those wood samples used in state-of-the-art

works are fresh and then with still strong and intense

aroma. By contrast, we use timber samples that, al-

though are drier and have a weaker aroma, they rep-

resent in a better way the actual scenarios of illegal

logging. As a result, error rates are higher than re-

sults reported in the literature (Cordeiro et al., 2016;

Wilson, 2012; Kalaw and Sevilla, 2018).

Finally, there is no information about which are

those appropriate chemical sensors for this speciﬁc

application. Therefore, it is suggested, as future work,

to perform an analysis of volatile compounds that

compounds the aromas of the different wood/timber

species.

ACKNOWLEDGEMENTS

We are grateful to professor Iv

an Porras for his valu-

able comments in early stages of the experiment. This

work was ﬁnanced in part by the research project

Plataforma IoT para el desarrollo de servicios in-

teligentes de apoyo al monitoreo ambiental–1971,

VIE- Universidad Industrial de Santander; and the

other part by E3T-Universidad Industrial de San-

tander, http://www.uis.edu.co.

REFERENCES

Akbar, M. A., Ait Si Ali, A., Amira, A., Bensaali, F., Be-

nammar, M., Hassan, M., and Bermak, A. (2016).

An empirical study for pca- and lda-based feature re-

duction for gas identiﬁcation. IEEE Sensors Journal,

16(14):5734–5746.

Bishop, C. M. (2006). Pattern recognition and machine

learning. springer.

Cabral, E. C., Simas, R. C., Santos, V. G., Queiroga, C. L.,

Da Cunha, V. S., De S

a, G. F., Daroda, R. J., and

Eberlin, M. N. (2012). Wood typiﬁcation by Venturi

easy ambient sonic spray ionization mass spectrome-

try: The case of the endangered Mahogany tree. J.

Mass Spectrom., 47(1):1–6.

Capelli, L., Sironi, S., and Del Rosso, R. (2014). Elec-

tronic Noses for Environmental Monitoring Applica-

tions. Sensors, 14(11):19979–20007.

Carmel, L., Levy, S., Lancet, D., and Harel, D. (2003).

A feature extraction method for chemical sensors in

electronic noses. Sensors and Actuators B: Chemical,

93(1):67 – 76. Proceedings of the Ninth International

Meeting on Chemical Sensors.

Cordeiro, J. R., Li, R. W. C., Takahashi,

E. S., Rehder, G. P.,

Ceccantini, G., and Gruber, J. (2016). Wood iden-

tiﬁcation by a portable low-cost polymer-based elec-

tronic nose. RSC Adv., 6(111):109945–109949.

Dickson, A., Nanayakkara, B., Sellier, D., Meason, D.,

Donaldson, L., and Brownlie, R. (2017). Fluorescence

imaging of cambial zones to study wood formation in

Pinus radiata D. Don. Trees - Structure and Function,

31(2):479–490.

Doddington, G. R., Przybocki, M. A., Martin, A. F., and

Reynolds, D. A. (2000). The NIST speaker recog-

nition evaluation - overview, methodology, systems,

results, perspective. Speech Communication, 31(2-

3):225–254.

Fedele, R., Galbally, I. E., Porter, N., and Weeks, I. A.

(2007). Biogenic VOC emissions from fresh leaf

ICPRAM 2020 - 9th International Conference on Pattern Recognition Applications and Methods

340

mulch and wood chips of Grevillea robusta (Aus-

tralian Silky Oak). Atmos. Environ., 41(38):8736–

8746.

Figaro (2005). General information for TGS sensors.

Garneau, F. X., Riedl, B., Hobbs, S., Pichette, A., and

Gagnon, H. (2004). The use of sensor array technol-

ogy for rapid differentiation of the sapwood and heart-

wood of Eastern Canadian spruce, ﬁr and pine. Holz

als Roh - und Werkstoff, 62(6):470–473.

Ghasemi-Varnamkhasti, M., Mohammad-Razdari, A.,

Yooseﬁan, S. H., Izadi, Z., and Rabiei, G. (2019).

Selection of an optimized metal oxide semiconduc-

tor sensor (mos) array for freshness characterization

of strawberry in polymer packages using response sur-

face method (rsm). Postharvest Biology and Technol-

ogy, 151:53 – 60.

Guo, L., Yang, Z., and Dou, X. (2017). Artiﬁcial Olfactory

System for Trace Identiﬁcation of Explosive Vapors

Realized by Optoelectronic Schottky Sensing. Adv.

Mater., 29(5):1–8.

Hanssen, F., Wischnewski, N., Moreth, U., and Magel,

E. A. (2011). Molecular identiﬁcation of Fitzroya cu-

pressoides, Sequoia sempervirens, and Thuja plicata

wood using taxon-speciﬁc rDNA-ITS primers. IAWA

J., 32(2):273–284.

Jolliffe, I. (2011). Principal component analysis. Springer.

Kalaw, J. M. and Sevilla, F. B. (2018). Discrimi-

nation of wood species based on a carbon nan-

otube/polymer composite chemiresistor array. Holz-

forschung, 72(3):215–223.

Kinnunen, T. and Li, H. (2010). An overview of text-

independent speaker recognition: From features to su-

pervectors. Speech communication, 52(1):12–40.

Martin, A., Doddington, G., Kamm, T., Ordowski, M., and

Przybocki, M. (1997). The det curve in assessment

of detection task performance. Technical report, Na-

tional Inst of Standards and Technology Gaithersburg

MD.

uller, K., Haferkorn, S., Grabmer, W., Wisthaler, A.,

Hansel, A., Kreuzwieser, J., Cojocariu, C., Rennen-

berg, H., and Herrmann, H. (2006). Biogenic carbonyl

compounds within and above a coniferous forest in

Germany. Atmos. Environ., 40:81–91.

Rana, R., M

uller, G., Naumann, A., and Polle, A. (2008).

FTIR spectroscopy in combination with principal

component analysis or cluster analysis as a tool to dis-

tinguish beech (Fagus sylvatica L.) trees grown at dif-

ferent sites. Holzforschung, 62(5):530–538.

Reynolds, D. A., Quatieri, T. F., and Dunn, R. B. (2000).

Speaker veriﬁcation using adapted gaussian mixture

models. Digital signal processing, 10(1-3):19–41.

Reynolds, D. A. and Rose, R. C. (1995). Robust

text-independent speaker identiﬁcation using gaussian

mixture speaker models. IEEE transactions on speech

and audio processing, 3(1):72–83.

Rinne, H. J. I., Guenther, A. B., Greenberg, J. P., and

Harley, P. C. (2002). Isoprene and monoterpene ﬂuxes

measured above Amazonian rainforest and their de-

pendence on light and temperature. Atmos. Environ.,

36(14):2421–2426.

Rodriguez-Lujan, I., Fonollosa, J., Vergara, A., Homer, M.,

and Huerta, R. (2014). On the calibration of sensor

arrays for pattern recognition using the minimal num-

ber of experiments. Chemometrics Intellig. Lab. Syst.,

130:123–134.

Ruiz Jim

enez, L. F. (2018). Detecci

on de los insec-

tos de la subfamilia Triatominae basado en narices

electr

onicas. Technical report, Universidad Industrial

de Santander.

Santos, J. P. and Lozano, J. (2015). Real time detec-

tion of beer defects with a hand held electronic nose.

Proceedings of the 2015 10th Spanish Conference on

Electron Devices, CDE 2015, pages 1–4.

Shi, H., Zhang, M., and Adhikari, B. (2017). Advances of

electronic nose and its application in fresh foods: A

review. Crit. Rev. Food Sci. Nutr., 8398:1–11.

Wheeler, E. A. and Baas, P. (1998). Wood identiﬁcation-a

review. IAWA journal, 19(3):241–264.

Wilson, A. D. (2012). Application of a Conductive Poly-

mer Electronic-Nose Device to Identify Aged Woody

Samples. The Third International Conference on Sen-

sor Device Technologies and Applications, pages 77–

82.

Wilson, A. D., Lester, D. G., and Oberle, C. S. (2005).

Application of conductive polymer analysis for wood

and woody plant identiﬁcations. For. Ecol. Manage.,

209(3):207–224.

Yan, J., Guo, X., Duan, S., Jia, P., Wang, L., Peng, C., and

Zhang, S. (2015). Electronic nose feature extraction

methods: A review. Sensors, 15(11):27804–27831.

Yu, M., Liu, K., Zhou, L., Zhao, L., and Liu, S. (2016).

Testing three proposed DNA barcodes for the wood

identiﬁcation of Dalbergia odorifera T. Chen and Dal-

bergia tonkinensis Prain. Holzforschung, 70(2):127–

136.

A Low Cost Electronic Nose with a GMM-UBM Approach for Wood Species Veriﬁcation

341