Classification of Honeybee Infestation by Varroa Destructor using Gas

Sensor Array

Andrzej Szczurek

, Monika Maciejewska

, Beata Bąk

, Jakub Wilk

, Jerzy Wilde

and Maciej Siuda

Faculty of Environmental Engineering, Wroclaw University of Science and Technology, Wyb. Wyspiańskiego 27,

50-370 Wrocław, Poland

Apiculture Department, Warmia and Mazury University in Olsztyn, Sloneczna 48, 10-957 Olsztyn, Poland

Keywords: Semiconductor Gas Sensor, Beehive, Indoor Air, Varroosis, Measurement.

Abstract: Infestation of bee colony with Varroa destructor proceeds exponentially. It is important to detect the disease

at its very early stage. However, the distinction of later infestation stages is also practical. We proposed to

apply gas sensor array measurements of beehive air as the source of information which may be useful for this

kind of assessment. Honeybee infestation was classified into three categories: ‘low’, ‘medium’ and ‘high’,

two categories: ‘low’ and ‘medium to high’, and another two categories: ‘high’ and ‘medium to low’.

Responses of gas sensor array to beehive air were used as the input data of the classifier, which was trained

to distinguish the categories. The results of the analysis demonstrated that category ‘low’ was determined

most effectively, with an error rate of about 10%. Category ‘high’ was most difficult to determine. In this case

the lowest error rate was about 20%. Based on our analysis, the approach based on binary classification was

favoured and SVM outperformed ensemble of classification trees. It was found, that first several minutes of

gas sensors exposure to beehive air were sufficient to attain effective classification. The presented method of

varroosis determination, based on beehive air sensing with gas sensors is innovative and has high potential of

application in beekeeping.

1 INTRODUCTION

Bees are critically important for the environment and

to the economy. They play a vital role in the

environment by pollinating both wild flowers and

many crops. Honey bees also provide honey and other

apiculture products such as pollen, wax, propolis and

royal jelly. Unfortunately, the population of these

insects is decreasing at an alarming rate throughout the

world. This phenomenon is still poorly understood

(EPILOBEE, 2016). Probably, it is caused by the

combined effect of interrelated factors, e.g. shifting

flowering seasons due to climate change, reduced

floral diversity, use of pesticides, habitat loss, lack of

genetic diversity, insect parasites and harmful

microorganisms.

https://orcid.org/ 0000-0002-7486-4552

https://orcid.org/ 0000-0001-6472-7944

https://orcid.org/ 0000-0002-9383-5335

https://orcid.org/ 0000-0003-4692-5474

https://orcid.org/ 0000-0002-3137-7957

https://orcid.org/ 0000-0002-1903-8688

The best source of highly reliable information

about the condition of bee colony, events that may

require the beekeeper's action and environmental

conditions affecting the colony health is beehive

monitoring. It can be based on regular inspection or

measurements of the appropriate parameters

(Sperandio et al., 2019 ). The first approach requires a

great experience. It is time-intensive and subject to

observer error. Hence, the measurement strategy is

preferred. In practice, beehive monitoring is focused

on the continuous, automatic determination of

temperature, air humidity and gas content, analysis of

sound and vibration of a beehive, counting of outgoing

and incoming bees, video observation, weighing the

colony (Cecchi et al., 2019; Kviesis, 2015). The data

Szczurek, A., Maciejewska, M., B ˛ak, B., Wilk, J., Wilde, J. and Siuda, M.

Classiﬁcation of Honeybee Infestation by Varroa Destructor using Gas Sensor Array.

DOI: 10.5220/0009171100610068

In Proceedings of the 9th International Conference on Sensor Networks (SENSORNETS 2020), pages 61-68

ISBN: 978-989-758-403-9; ISSN: 2184-4380

can be provided in real time and used for individual bee

colony maintenance.

Substantial information about honeybee colony is

included in the chemical composition of air inside the

hive. Usually, this gas is a mixture of compounds

emitted by the bees themselves (e.g. pheromones, other

chemicals released to repel pests and predators,

metabolites, etc.), substances originating from hive

stores (e.g. honey, nectar, larvae, sealed brood,

beeswax, pollen, beebread and propolis), and volatile

compounds from hives construction materials (wood,

paint, plastic, etc.). The beehive atmosphere also

contains compounds which come from vehicles, farms,

industries, and households located in the hive vicinity.

The combination of gaseous mixture inside the

hive is unique. Sometimes bee diseases can influence

the indoor air of a hive. For example, foulbrood has a

characteristic odor, and experienced beekeepers, with

a good sense of smell, can detect the disease upon

opening a hive. It is known, that the notorious varroa

mites can change their surface chemicals to match the

development stage of their hosts. It is interesting to

know if, despite this, the information about infestation

is included in the chemical composition of the air

surrounding bees inside the hive. Potentially, changes

of the chemical properties of the indoor air could be the

basis for detection of varroosis (Szczurek et al. 2019a;

Szczurek et al. 2019b). This most destructive disease

of honey bees worldwide is caused by a Varroa

destructor (V.d.).

Generally speaking, the determination of the

chemical indicators of varroosis can be based on the

detection of specific volatile chemicals, qualitative and

quantitative gas analysis and qualitative classification

of indoor air. Today, there are a number of well-

established methods which are capable of detecting the

specific chemical species or analysing the complex

gaseous mixture. They offer very good detection limit,

accuracy, sensitivity and repeatability. Unfortunately,

the available methods and instruments are expensive

and require trained, experienced personnel. In practice,

they are beyond the reach of average users -

beekeepers.

In this situation, the measurement instruments

based on gas sensors offer wider usefulness and

applicability. Chemiresistors are especially promising

in this field of application (Yunusa et al., 2014). These

devices present high sensitivity, detection at the level

of ppm, small sizes, low cost, simplicity of their use.

The serious shortcomings of the semiconductor gas

sensors is poor selectivity, resulting from the sensing

mechanism. For that reason, it is impossible to detect

individual chemical species using a single

semiconductor sensor. The measurement potential of

devices based on chemiresistors may be improved by

the application of the multi-sensor array, the

appropriate operation mode, signal processing and data

analysis. The instruments established on this idea are

particularly useful for the pattern recognition.

The aim of this study is to show that the

measurement instrument consisting of the sensor array

and the appropriate data classification module allows

to detect varroosis. The term detection in this work is

understood as the action of accessing information

about the rate of infestation. The effective detection

can include the determination of several levels of

infestation. We expected that the accuracy and

sensitivity of the detection process is strongly

influenced by the number of the assumed categories.

The determination of this relationship can be of major

importance in respect of the practical application of

sensor device as bee disease detector.

The main advantages of the presented method of

Varroa destructor classification based on gas sensing

are related to its cost-effectiveness, availability and low

detection limit. Continuous measurement may be

accomplished and the measurement data is provided in

real time. On-site detection can be performed. These

features are the good basis for establishing the honey

bee dieses monitoring system.

2 EXPARIMENTAL PART

2.1 Gas Sensor Device

Prototype Multisensor detector of air quality was used

in the study. The construction was developed in the

Laboratory of Sensor Technique and Indoor Air

Quality Studies at Wroclaw University of Science and

Technology, Poland. This autonomous,

multifunctional and programmable device is based on

gas sensors. It allows for continuous measurements of

gas samples and remote access to the recorded data.

The prototype was designed to operate in field

conditions. For this purpose, it was fitted with solar

panels, battery and the cover, which protects against

the meteorological conditions. The general view of the

instrument is presented in Figure 1.

The instrument was composed of several

functional modules: 1) multichannel recorder of gas

sensor signals MCA-8, 2) communication controller

Beecom, 3) charging regulator for solar panel, Steca

Solsum 6.6 , 4) gel battery, HZY EV12-33, with the

nominal power 36 Ah and voltage 12V and battery

level indicator, 5) photovoltaic solar panel, CL050-

12P, with the nominal power 50W, 7) AC adapters and

6) casing.

SENSORNETS 2020 - 9th International Conference on Sensor Networks

The major functional unit of the device was the

multichannel recorder of gas sensor signals MCA-8.

It included the following gas sensors TGS832,

TGS2602, TGS823, TGS826, TGS2603, and

TGS2600. They were mounted in individual sensor

chambers, made of aluminum. Gas sensors heaters

temperature was stabilized.

The device was dedicated to operate continuously

and perform measurements in the dynamic mode. A

peristaltic pump was mounted inside in order to

enforce the gas flow. The instrument was fitted with

8 gas inlet ports, which could be individually

connected to gas sensors chambers by means of a set

of valves. This solution allows for an intermittent gas

sampling from 8 locations.

Figure 1: General view of the instrument based on gas

sensors.

As default, the measurement data is recorded on

the instrument’s SD card, with the temporal

resolution of 1s. Optionally, the remote data transfer

could be realized using GSM (5s resolution). The

operation of the gas sensor device is programmable.

The user has to define the following parameters:

duration of gas intake through individual inlet ports,

pump operation rate as well as the power of gas

sensors heaters. The program is executed from SD

card. The instrument may be also operated in an

interactive mode using a PC based software.

Three options of powering the device are

available: mains power supply, battery and

photovoltaic solar panel. The last two solutions were

aimed to secure the autonomous operation of the

device in field conditions.

2.2 Field Experiments

Fifteen honey bee colonies were chosen for the

experiment. They belonged to three groups called A,

B and C. Each group included five colonies. Groups

differed in respect of the degree of Varroa destructor

infestation. The infestation rates of individual

colonies are shown in Table 1, Table 2 and Table 3.

The Varroa destructor infestation rates of honey

bee colonies were determined using a flotation

method. It involves shaking a sample of dead bees

with a detergent or alcohol and then rinsing them on

a sieve (COLOSS BEEBOOK, 2013; Fries et al.,

1991). The infestation rate is the number of mites

found in a sample of bees, divided by the number of

bees and expressed as the percentage.

Table 1: Honey bee colonies which belonged to group A.

Colony V.d. infestation rate

A1 0.0

A2 0.6

A3 0.0

A4 0.3

A5 0.2

Table 2: Honey bee colonies which belonged to group B.

Colony V.d. infestation rate

B1 4.9

B2 4.7

B3 4.4

B4 3.8

B5 4.3

Table 3: Honey bee colonies which belonged to group C.

Colony V.d. infestation rate

C1 60.3

C2 52.0

C3 11.0

C4 11.5

C5 13.0

The air of beehives occupied by bees was

measured using gas senor device.

The measurement experiment lasted five days.

Each day, three bee colonies, were investigated, one

from group A, B and C. A single measurement of a

bee colony consisted of two phases: 1) the exposure

of gas sensors to beehive air (600 s), 2) the exposure

of gas sensors to the regeneration air (900 s). The

measurements of three individual colonies were

performed in sequence. The sequence was repeated,

Classiﬁcation of Honeybee Infestation by Varroa Destructor using Gas Sensor Array

therfore multiple measurements were done for each

bee colony.

During measurements, gas sensor device was

connected to beehives by means of polyethylene

tubing. One inlet port was used to deliver the air

sampled from one beehive. The gas sampling points

were located inside hives, in their central, upper parts,

between brood combs. In this location, the bee colony

infestation should be most strongly reflected in the

quality of beehive air, because the mite proliferates

on the brood. One additional inlet port of gas sensor

device was dedicated to the delivery of ambient air

for sensors regeneration. Dedicated filter, filled with

charcoal allowed for air preparation. Inlet ports of the

device were protected by particle filters.

The experiment was run in field condidtions.

3 DATA ANALYSIS

3.1 Varroa Destructor Infestation

Category

V.d. infestation rate of bee

colony [%]

High >6

Medium to low 0-6

to 0.6%, see Table 1. The category ‘medium’ was

represented by colonies group B. In this group the

V.d. infestation rate was from 3.8% to 4.9%, see

Table 2. The category ‘high’ was represented by bee

colonies group C. Here, the V.d. infestation rate was

from 11% to 60.3%, see Table 3.

The problem of recognition of three categories of

infestation was represented by a three-class

classification task.

Second approach consisted in distinguishing two

categories of V.d. infestation , which were ‘low’ and

‘medium to high’, as shown in Table 5. This approach

could be used to filter out bee colonies which are not

infested or slightly infested, perhaps not yet requiring

treatment, from all other infested colonies.

The third approach also consisted in

distinguishing two categories of V.d. infestation.

However, the considered categories were ‘high’ and

‘medium to low’, as shown in Table 6. This approach

could be used to detect bee colonies which are

severely infested, and should be subject to a radical

treatment, from other less infested or even healthy

colonies.

The problem of recognition of two categories of

infestation was represented by a binary classification

task. The distinction of two categories is less

attractive for the beekeeper. However, this approach

is likely to offer a trade off in terms of smaller

classification error. Two-class problems may be

solved using wide range of classifiers, which are not

available in case of multiclass classification.

3.2 Classification

Classification was based on responses of all sensors,

which were elements of gas sensor array. In order to

form a feature vector, a 3 min long fragment was

extracted from the signal of each gas sensor. It

consisted of 180 responses recorded one after another

with temporal resolution of 1s. Fragments of signals

of all sensors were combined to form one feature

vector.

Several feature vectors were considered in this

work as the basis of classification. They included

fragments associated with first, second, third, fourth

etc. three minutes of gas sensor array exposure to the

test gas – beehive air.

SENSORNETS 2020 - 9th International Conference on Sensor Networks

Two classifiers were applied. Ensemble of

classification trees (Ren et al., 2016) and support

vecor machine (SVM) (Nalepa and Kawulok, 2019).

The first classifier was utilised for solving three-class

as well as two-class classification problems. SVM

was applied for binary problems solving, exclusively.

The performance of classification was evaluated

based on confusion matrices, as shown in Table 7 and

Table 8.

Table 7: Confusion matrix for two-class problem.

Predicted cat. 1 Predicted cat. 2

True cat. 1 n

1,1

1,2

True cat. 2 n

2,1

2,2

Regarding two-class problem (see Table 7), the

rate of correct classification (TC) of data representing

category 1 was called TC1 rate and it was given by

eq. 1.

TC1 rate = n

1,1

/(n

1,1

1,2

) (1)

TC2 rate was determined analogically.

Table 8: Confusion matrix for three-class problem.

Predicted

cat. 1

Predicted

cat. 2

Predicted

cat. 3

True cat. 1 n

1,1

1,2

1,3

True cat. 2 n

2,1

2,2

2,3

True cat. 3 n

3,1

3,2

3,3

In case of three-class problem (see Table 8), the

rate of correct classification of data representing

category 1 was given by eq. 2.

TC1 rate = n

1,1

/(n

1,1

1,2

1,3

) (2)

TC2 rate and TC3 rates were determined

analogically.

Classification models were validated using ten-

folds cross-validation procedure. It was repeated

fifteen times for each classifier, when using a

particular feature vector as input. The results of

repeated cross-validations were averaged and

standard deviation was computed. Following,

confusion matrices were prepared which included the

averaged results as well as the information about their

spread.

4 RESULTS

4.1 Three Categories of Varroa

Destructor Infestation: ‘Low’,

‘Medium’ and ‘High’

Three categories of V.d. infestation: ‘low’, ‘medium’

and ‘high’ were recognized using one classifier,

ensemble of classification trees. The classification

performance was examined with respect to different

fragments of gas sensors signals, utilised as the basis

of classification. Figure 2 presents the results of

classification in terms of True Category rates for each

category.

As shown in Figure 2, three considered infestation

categories were distinguished with various

efficiencies. The rate of correct classifications was

the best in case of category ‘low’ (on average, TC1

rate was from 80% to 87%). Smaller rates were

associated with category ‘medium’ (on average TC2

rate was from 72% to 76%) and the worst results were

attained in case of category ‘high’ (on average, TC3

rate was from 66% to 74%).

Similar results were obtained when using

different fragments of gas sensors signals as the bass

of classification. TC rates varied as a function of the

duration of gas sensors exposure to beehive air, but

no clear relationship was observed between the two.

Based on the obtained results (see Figure 2), first

three minutes of gas sensors exposure could be

considered sufficient for collecting the informative

measurement data, useful for classification.

Figure 2: Results of classification for three categories of

V.d. infestation: ‘low’, ‘medium’ and ‘high’. Ensemble of

classification trees was applied and various fragments of

gas sensors signals recorded during exposure to beehive air

were utilised as the classifier input.

Table 9 presents a confusion matrix for

classification based on first three minutes of gas

Classiﬁcation of Honeybee Infestation by Varroa Destructor using Gas Sensor Array

sensor array exposure to beehive air. As shown,

misclassified data, truly belonging to class ‘low’ were

mostly allocated to class ‘medium’ (12.1%) and the

remaining 7.9% was assigned to class ‘high’.

Majority of misclassified items, truly belonging to

class ‘medium’ was recognized as members of

category ‘high’ (19.8%) and only 6.2% of them were

allocated to class ‘low’. In case of category ‘high’,

also 24.3% of misclassified data were allocated to

class ‘medium’ and only 7.6% were assigned to class

‘low’.

Table 9: Confusion matrix for recognition of three

categories of V.d. infestation: ‘low’, ‘medium’ and ‘high’.

Mean±standard deviation for 15 cross-validations are

shown. Ensemble of classification trees was applied. Input

data consisted of first three minutes of gas sensor array

responses to beehive air.

Predicted

‘low’

Predicted

‘medium’

Predicted

‘high’

True ‘low’ 80.0±3.1 12.1±2.3 7.9±2.3

True ‘medium’ 6.2±1.6 74±4.9 19.8±5.0

True ‘high’ 7.6±1.9 24.3±3.9 68.1±5.0

It should be noted that regarding extreme

categories ‘low’ and ‘high’, the structure of

misclassified items allocation was logical. Namely,

most of overlaps were between the directly

neighbouring classes, ‘low’ with ‘medium’ and ‘high’

with ‘medium’. In case of category ‘medium’, the

misclassified items mostly fell in the category ‘high’

and much less of them was recognized as members of

category ‘low’. This asymmetry indicates

considerable similarity of categories ‘medium’ and

‘high’, while category ‘low’ was more distinct than

the two.

4.2 Two Categories of Varroa

Destructor Infestation: ‘Low’ and

‘Medium to High’

Another approach consisted in determining two

categories of V.d. infestation, namely ‘low’ and

‘medium to high’. Classification was realised using

ensemble of classification trees and SVM. The results

are shown in Figure 3 and Figure 4, respectively.

From the comparison of TC rates obtained when

using ensemble of classification trees, infestation

categories ‘low’ and ‘medium to high’ were

distinguished more effectively (see Figure 3) than

categories ‘low’, ’medium’ and ‘high’(see Figure 2),.

TC rates associated with the recognition of categories

‘low’ and ‘medium to high’ were similar, at the level

of about 83%, see Figure 5. Still, a considerable

improvement could be achieved by applying another

classifier. In case of using SVM, TC rate was 93% for

category ‘low’ and 88% for category ‘medium to

high’, see Figure 4. This result shall be recognized as

very good.

Figure 3: Results of classification for two categories of V.d.

infestation: ‘low’ and ‘medium to high’. Ensemble of

classification trees was applied and various fragments of

gas sensors signals recorded during exposure to beehive air

were utilised as the classifier input.

Figure 4: Results of classification for two categories of V.d.

infestation: ‘low’ and ‘medium to high’. SVM was applied

and various fragments of gas sensors signals recorded

during exposure to beehive air were utilised as the classifier

input.

It has to be added that the ensemble of

classification trees was relatively insensitive to the

fragment of gas sensor signal utilised as the source of

input data. SVM favoured the information acquired

during early stages of gas sensors exposure.

SENSORNETS 2020 - 9th International Conference on Sensor Networks

4.3 Two Categories of Varroa

Destructor Infestation: ‘High’ and

‘Medium to Low’

Additionally, the distinction of V.d. infestation

categories ‘high’ and ‘medium to low’ was

considered. Classification was based on gas sensor

array measurements and it was realised using

ensemble of classification trees and SVM. The results

are shown in Figure 5 and Figure 6, respectively.

Figure 5: Results of classification for two categories of V.d.

infestation: ‘high’ and ‘medium to low’. Ensemble of

classification trees was applied and various fragments of

gas sensors signals recorded during exposure to beehive air

were utilised as the classifier input.

Figure 6: Results of classification when two categories of

V.d. infestation are distinguished: ‘high’ and ‘medium to

low’. SVM was applied and various fragments of gas

sensors signals recorded during exposure to beehive air

were utilised as the classifier input.

From the comparison of TC rates obtained when

using ensemble of classification trees, infestation

categories ‘high’ and ‘medium to low’ (see Figure 5)

were determined more effectively than categories

‘high’ and ‘medium’ (see Figure 2). However, the

attained improvement was not substantial. True

Category rate was, on average, 74% for recognition

of category ‘high’ and 80% for category ‘medium to

high’, see Figure 5. The change of classifier to SVM

resulted in the increase of the classification

performance indicators, up to the level of 80% and

84%, respectively (see Figure 6). In case of both

classifiers, late fragments of gas sensor array signals

were favoured as the sources of information.

5 DISCUSSION

Three approaches to classification of V.d. infestation

of bee colonies were compared in this work. They

consisted in the determination of:

1. three categories of infestation: ‘low’, ‘medium’

and ‘high’;

2. two categories of infestation: ‘low’ and ‘medium

to high’;

3. two categories of infestation: ‘high’ and

‘medium to low’.

The first approach was realised using ensemble of

classification trees. In case of the second and third

approach SVM was applied, additionally.

Ensemble of classification trees is applicable to

both binary and multi class problems. Nevertheless,

when using this method for binary problems (second

and third approach) better results were attained, as

compared with the multi class problem (first

approach). Regarding binary classification tasks,

further improvement was possible by applying SVM.

Clearly, the best discernible V.d. infestation

category was ‘low’, no matter if binary or three class

classification problem was formulated and solved. In

this case, the best attained TC1 rates were at the level

of 90%. This result indicates that conditions of no

infestation or very weak infestation are quite clearly

discernible from the conditions of medium or

advanced infestation, based on gas sensor array

responses.

The detection of V.d. infestation category ’high’,

was most difficult. The analysis of assignment of

misclassified data, truly belonging to this category

indicated a considerable overlap with the category

‘medium’ and vice versa. When realised in the

framework of binary classification task, the detection

of ‘high’ infestation rate was more effective. In

particular, the use of SVM allowed for achieving

TC1 rate about 80%.

The classification performance was examined for

several fragments of gas sensors signals utilised as the

sources of input data for the classifier. Clearly, the

Classiﬁcation of Honeybee Infestation by Varroa Destructor using Gas Sensor Array

duration of gas sensors exposure to beehive air had an

influence on the recognition of infestation categories.

However, we have not identified any fragment which

could be definitely preferred. One could notice, that

the measurement data collected during first three

minutes of gas sensors exposure to beehive air is a

reasonable source of information about the

infestation. This result justifies the measurement

procedure which includes a relatively short period of

gas sensors exposure to beehive air. This observation

is highly beneficial from practical point of view.

6 CONCLUSIONS

The paper was dedicated to the recognition of several

categories of bee colonies infestation by Varroa

destructor, based on responses of gas sensor array to

beehive air.

The results of the analysis demonstrated that first

several minutes of gas sensors exposure to beehive air

were sufficient to attain effective classification.

Category representing ‘low’ infestation was

determined most effectively, with an error rate of

about 10%. Category ‘high’ was most difficult to

determine. In this case the lowest error rate was about

20%. The approach based on binary classification

granted higher performance as compared with three

class classification. SVM outperformed ensemble of

classification trees.

ACKNOWLEDGEMENTS

This work was supported by the National Centre for

Research and Development under the grant nr

BIOSTRATEG3/343779/10/NCBR/2017

“Developing innovative, intelligent tools to

monitoring the occurrence of malignant foulbrood

and elevated levels of infestation with Varroa

destructor in honey bee colonies.”

REFERENCES

Cecchi, S., Terenzi, A., Orcioni, S., Spinsante, S., Mariani

Primiani, V., Moglie, F., Ruschioni, S., Mattei, C.,

Riolo, P., Isidoro, N., 2019. Multi-sensor platform for

real time measurements of honey bee hive parameters,

IOP Conf. Series: Earth and Environmental Science

275 , 012016IOP Publishing.

COLOSS BEEBOOK, 2013. Volume II: Standard methods

for Apis mellifera pest and pathogen research, Ed. by.

V. Dietemann, J. D. Ellis, P. Neumann, IBRA,.

EPILOBEE Consortium; Chauzat, M.-P., Jacques,

A.; Laurent, M., Bougeard, S., Hendrikx, P., Ribière-

Chabert, M., 2016. Risk indicators affecting honeybee

colony survival in Europe: one year of surveillance,

Apidologie 47(3), 348–378.

Fries, I., Aarhus, A., Hansen, H., Korpela, S., 1991.

Comparisons of diagnostic methods for detection of

Varroa jacobsoni in honey bee (Apis mellifera) colonies

at low infestation levels. Experimental and Applied

Acarology 10, 279–287.

Kviesis, A., Zacepins, A., Durgun, M., Tekin, S., 2015.

Application of wireless sensor networks in precision

apiculture. Engineering for Rural Development 2015,

Jelgava, 20–22.

Nalepa J., Kawulok, M., 2019. Selecting training sets for

support vector machines: a review. Artificial

Intelligence Review 52(2), 857–900.

Ren, Y., Zhang, L., Suganthan, P., 2016. Ensemble

classification and regression – recent developments

applications and future directions, IEEE Computational

Intelligence Magazine 11 (1), 41-53.

Sperandio, G., Simonetto, A., Carnesecchi, E., Costa, C.,

Hatjina, F., Tosig, S., Gilioli, G. 2019. Beekeeping and

honey bee colony health: A review and

conceptualization of beekeeping management practices

implemented in Europe, Science of The Total

Environment 696, 133795.

Szczurek, A., Maciejewska, M., Bąk, B., Wilde, J., Siuda,

M., 2019a. Semiconductor gas sensor as a detector of

Varroa destructor infestation of honey bee colonies –

Statistical evaluation, Computers and Electronics in

Agriculture 162, 405–411.

Szczurek, A., Maciejewska, M., Bąk, B., Wilk, J., Wilde,

J., Siuda, M., 2019b. Detection of honeybee disease:

varrosis using a semiconductor gas sensor array.

SENSORNETS 2019, SciTePress, 58-66.

Yunusa, Z., Hamidon, M.N., Kaiser, A., Awang, Z., 2014.

Gas sensors: a review. Sensors and Transducers Jurnal

168 (4), 61–75.

SENSORNETS 2020 - 9th International Conference on Sensor Networks