Computers and Electronics in Agriculture 237 (2025) 110508

Contents lists available at ScienceDirect

Computers and Electronics in Agriculture

journal homepage: www.elsevier.com/locate/compag

Original papers

Determination of flowering stage based on artificial intelligence and the daily

weight of bee hives
Andrés Gersnoviez a , Francisco J. Rodriguez-Lozano a ,∗, María Brox a ,∗,
José Moreno-Carbonell b , Manuel Ortiz-Lopez a , José M. Flores c
a
Department of Electronic and Computer Engineering, Universidad de Córdoba, Córdoba, 14071, Spain
b Veterinary Management Studies (VMS), Lleida, 25191, Spain
c Department of Zoology, Universidad de Córdoba, Córdoba, 14071, Spain

ARTICLE INFO ABSTRACT

Keywords: Honey bee plays a very important role in pollination and is essential for the balance of terrestrial ecosystems
Classifier system and in the pollination of important crops. The success of honey bee hives and beekeeping depends on the
Flowering flowering period, and good hive management during this period is essential for beekeepers. The use of new
Honey bees
technologies in beekeeping can help this farming activity enormously. Based on a monitoring system of several
Machine learning
hives located in the south of Spain, this work presents a study of the data obtained to find out if there is a
Precision apiculture
relationship between these data and the flowering stage of the hives. In this study, it is determined that
the evolution of the weight of the hive throughout the day is crucial to determine the flowering stage. By
testing the behavior of several machine learning algorithms, a highly efficient classifier is obtained, capable
of determining which stage of flowering the hives are in. It is able not only to determine whether the hives
are before, during or after flowering, but also to distinguish between an initial and final stage of flowering.
This is important because it can enable beekeepers to effectively plan apiary visits, hive maintenance work
and honey harvesting, making beekeeping more profitable.

1. Introduction goes to different locations, which greatly complicates the monitoring of

the hives, forcing a large number of journeys, with significant time and
The honey bee Apis mellifera L. is the basis of western beekeep- economic costs. So being able to know the state of the hives remotely,
ing (Eimanifar et al., 2020). Bees and beekeepers make up a very from the farm base or from anywhere, is a desire for these professionals.
particular livestock sector, since bees, in addition to producing honey, And new technologies are beginning to provide the tools to make this
pollen, royal jelly and wax, among other things, are responsible for
possible (Zacepins et al., 2015).
the pollination of a large number of plant species, including some
Remote monitoring systems already make it possible to know the
crops that are essential to our society (Prodanović et al., 2024; Khalifa
weight of the hives or the humidity and temperature conditions in the
et al., 2021). Unfortunately, these bees suffer multiple threats, and
practically all bee hives are kept by beekeepers, with hardly any wild microclimate of the hive, or the sounds that the bees are generating,
swarms (Genersch, 2010). Beekeeping is a different kind of animal among others (Zacepins et al., 2015; Hadjur et al., 2022; Quiles-Latorre
husbandry, not tied to a defined and permanent plot of land. This is et al., 2024; McMinn-Sauder et al., 2024). And what is more important,
because bees make long flights, in search of food, beyond the strict their evolution. And all from anywhere, allowing the beekeeper to plan
location of the hives (Pahl et al., 2011). But also because it is common his visits to the hives at the right times and at the lowest costs (Ochoa
for professional beekeepers to move their hives in search of flowering. et al., 2019b; Alleri et al., 2023; Hadjur et al., 2022; Kiromitis et al.,
Sometimes many hundreds or thousands of kilometers away, in what 2022). A good example of this is the exploitation of the flowering
is known as transhumance (FAO, 2008; FAO et al., 2021). This poses period, since professional transhumant beekeepers do not have to keep
the problem of monitoring the hives that have been moved. Even more their hives in the settlements beyond the end of the flowering period
so when beekeepers gather the hives in small groups (no more than 80 they are using. It is then time to collect the honey harvest and move the
hives each is recommended) (Komasilova et al., 2021) and each group

∗ Corresponding authors.
E-mail addresses: [email protected] (A. Gersnoviez), [email protected] (F.J. Rodriguez-Lozano), [email protected] (M. Brox),
[email protected] (J. Moreno-Carbonell), [email protected] (M. Ortiz-Lopez), [email protected] (J.M. Flores).

https://doi.org/10.1016/j.compag.2025.110508
Received 26 February 2025; Received in revised form 18 April 2025; Accepted 4 May 2025
Available online 26 May 2025
0168-1699/© 2025 The Authors. Published by Elsevier B.V. This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/).
A. Gersnoviez et al. Computers and Electronics in Agriculture 237 (2025) 110508

hives to other flowering periods. If the hives are removed earlier, they bee colony weight is a key parameter to knowing the health and
may be losing part of that harvest, but if they are kept beyond the end state of the colony since changes in weight can provide important
of the flowering the bees may start to eat the stored honey, reducing information about this (Kviesis et al., 2020b; Terenzi et al., 2019;
the harvest as well (Albert Raj et al., 2022). Czekońska et al., 2023; Zacepins et al., 2017b; Meikle et al., 2016;
Nevertheless the task to accurately predict the end of the flowering Flores et al., 2019; Arias-Calluari et al., 2023; Meikle et al., 2008,
period by beekeepers is not trivial. There are studies that analyze flow- 2018; Bayir and Albayrak, 2016). As an example, Kviesis et al. (2020b)
ering with pollen traps (Liolios et al., 2023). However, these methods propose a development where the temperature and weight of a bee
can disrupt the work of the bees in the hives studied, because pollen is colony are measured using two temperature sensors and a load cell that
removed at the entrance to the hive. The alternative is therefore visual supports measurements of up to 200 kg. All the data provided by the
inspection by the beekeeper. However, determining the flowering stage sensors are collected by a Raspberry Pi and sent to a remote data store.
is guided by the experience of beekeepers and work organization. Raspberry Pi is also used by Terenzi et al. (2019), where a system is
Without these two fundamental factors in a beekeeper, but this can presented which, in addition to acquiring sound, temperature, humidity
lead them to visit the apiary early, can lead them to visit the apiary and CO2 data, also allows measurement of the weight using four load
prematurely and decide whether to remove the honey, or take the hives cells connected in a Wheatstone bridge and suitably placed to provide
to other flowering when the bloom may not yet be over, wasting the air flow at the bottom of the beehive. Weight and CO2 measurements
remaining flowering. In other occasions, if they delay the visit and combined with sound variation have been used to identify the swarm-
the flowering has already finished some time ago, the bees can be ing event. Czekońska et al. (2023) analyze how weather conditions,
consuming the stored honey, or mixing it with the one coming from landscape and forest structure influence the amount of food collected
other new flowering, which is a limitation when they want to produce by honey bees, using variations in hive weight as an indicator. Another
honeys composed by monofloral pollen (more valued). approach for bee colony weight measurement is proposed by Zacepins
Therefore, in order to understand the evolution of flowering, dif- et al. (2017b), which describes a system equipped with a temperature
ferent researches have used weight changes in hives (Zacepins et al., sensor, an RF transceiver, a battery, a set of photovoltaic cells, a charge
2015; McMinn-Sauder et al., 2024). In Flores et al. (2019), the authors controller, four strain gauge load cells for weight measurement, and
studied the effect of climate change on the exploitation of flowering by GSM/GPRS modules for communication with a remote server, Weight
bees on the basis of total daily weight increases in the hives, showing data, along with temperature and battery level measurements, are
a phase of significant daily increases during flowering and a phase transmitted to the cloud every 10 min. The relationship of the weight
with no representative weight changes when flowering had come to of a hive with important risks for honey bees such as the impact of
an end. But, from the perspective of the beekeeper, it is much more pesticides or climate change is studied by Meikle et al. (2016) and
interesting to foresee when that time is coming and to schedule his by Flores et al. (2019), respectively. Arias-Calluari et al. (2023) use
work than to find out that the flowering has already finished several daily hive weight fluctuations to develop a mathematical model that
days later (Kotovs and Zacepins, 2023). In this sense, the application estimates the number of bees engaged in foraging and the duration
of artificial intelligence, applied to the data being supplied by the hives, of their foraging activity. This model serves as a valuable tool for
can be a useful working tool (Turyagyenda et al., 2025; Astuti et al., beekeepers to monitor colony health and productivity. Meikle et al.
2024). (2008) record hourly weight data from four honey bee hives using
For this reason, this article is based on a database obtained by a balances which have a maximum capacity of 100 kg, operate with an
bee monitoring system located in the south of Spain, in order to find a accuracy of ±10 g and work at a temperature range from −30 ◦ C to
correlation between the measurements obtained and the behavior of the 70 ◦ C. The devices were powered by a solar panel and connected to
bees during the different stages of flowering. And, within the distinction 12-bit dataloggers. Overall, the weighing system has an accuracy of
between the pre-flowering, flowering and post-flowering stages, to be approximately ±30 g. Meikle et al. (2018) use continuous hive weight
able to distinguish, in addition, whether the flowering stage is in an monitoring to assess the impact of environmental factors on honey bee
initial or final stage. The latter is very important because, if it is known colonies. Data are analyzed using piecewise regression to determine the
that the flowering is coming to an end, it is time to alert the beekeeper start and end of the colony’s daily activity cycle, nocturnal variations
to harvest the honey before the bees start consuming it. in weight, the decrease in weight due to foragers leaving and the
Once this correlation has been obtained, the next step will be to corresponding increase when they return. The authors explain how
use different artificial intelligence algorithms to achieve an efficient daily weight monitoring can be a useful tool for detecting changes in
classifier system, capable of determining which stage of flowering the colony activity related to nutritional stress. Bayir and Albayrak (2016)
hives are in from the measurements taken on the same day. present a WSN where the weight of the hive as well as the humidity and
The article is organized as follows: Section 2 is dedicated to com- temperature of the environment are continuously monitored during the
pile different works in the literature focusing on bee monitoring; The flowering period from June to July. The real-time data can be accessed
monitoring system used to acquire the measurements is specified in via a web platform, and the length of the flowering period is identified
Section 3; Section 4 describes and analyzes the dataset obtained, as well from the variations in the hive weight. The WSN consists of two sensor
as the correlations discovered; Section 5 focuses on the design of the nodes and a coordinator node. Communication between the nodes is
classifier system, where different artificial intelligence algorithms are performed via ZigBee, while data transmission to the web is carried
used, as well as illustrating the results obtained by them; the article out via a Simcard. Each sensor node uses an Arduino Mega and has
ends in Section 6 where the conclusions are presented. a solar panel, a charge controller, and a battery for power supply.
The sensor node records humidity, temperature and weight data and
2. Related works sends the information to the coordinator node. To measure humidity
and temperature, an SHT11 sensor connected to an XBee module is
Currently there are numerous researches regarding honey bee mon- used. In addition, a load cell with a measuring capacity between 0 and
itoring to obtain a knowledge of the processes developed in a bee 100 kg is placed under each hive to continuously record changes in
colony (McMinn-Sauder et al., 2024; Sharif et al., 2022; Gil-Lebrero weight. In contrast to the sensor nodes, the coordinator node integrates
et al., 2016; Zacepins et al., 2017a; Cejrowski et al., 2019; Ochoa a GSM/GPRS module for transmitting the collected data to the internet.
et al., 2019a; Imoize et al., 2020; Henry et al., 2019; Meikle and On the other hand, with these measurement data, several works
Holst, 2015). Wireless sensor networks (WSN) are the main technology propose the use of fuzzy logic or neural networks to classify hive
used to develop this monitoring and the most widely measured data states (Kviesis et al., 2020a; Bassford and Painter, 2016; Edwards-
are humidity, temperature, sound, vibration and weight. Specifically, Murphy et al., 2016; Bayir and Albayrak, 2018; Kviesis and Zacepins,

2
A. Gersnoviez et al. Computers and Electronics in Agriculture 237 (2025) 110508

Fig. 1. Experimental beehives with the monitoring system.

2016; Degenfellner and Templ, 2024). For example, Kviesis et al. 1. Waspmote module: Equipped with the ATmega1281 microcon-
(2020a) present a proposal for determining the status of honeybee troller, this module is the core of the wireless node.
colonies based on the use of fuzzy logic and temperature analysis. Their 2. MicroSD card: Stores the collected data in case the node cannot
fuzzy inference system processes five input parameters and generates connect to the network or if there is a communication problem.
a result that classifies the status of the colony as normal, dead or 3. Real Time Clock (RTC): Used to clock the collected samples
extreme. Bassford and Painter (2016) also apply fuzzy logic in bioen- accurately and to wake up the microcontroller according to the
vironmental modeling to monitor honeybee colonies and predict their programmed frequency.
collapse. Similarly, Edwards-Murphy et al. (2016) employ a decision 4. Communication module connector: Allows connection of various
tree-based algorithm to identify ten different states in the hives. In addi- types of communication modules.
tion, an algorithm for local weather prediction is developed. The system 5. Microcontroller General Purpose Connector: Provides flexibility for
is based on an ATmega1281 microcontroller, which operates with low- connecting other devices or sensors.
power sleep modes to optimize its lifetime. In terms of communication, 6. Power control block: Manages system power.
ZigBee was used for the local network, while 3G technology was chosen
for long-distance transmission at the base station. In the work of Bayir Data collection is carried out by means of three SHT15 temperature
and Albayrak (2018) a neural network model with three inputs and and humidity sensors, located in three different layers of the hive (the
three outputs is designed to analyze the general condition of hives. innermost, an intermediate and the outermost), and a weighing scale
The input variables include humidity, weight and temperature, while (shown in Fig. 3(b)). The main characteristics of the latter are described
the outputs classify the hive condition into ‘‘good’’, ‘‘stable’’ or ‘‘bad’’ below:
situation. The model was implemented on an Arduino development
board and evaluated in real time, achieving an accuracy of 99.84%. • Construction: The scale has a metal frame and a 50 cm × 40 cm
Following the same line, a neural network was applied by Kviesis stainless steel plate.
and Zacepins (2016) to detect brood rearing and swarming stages in • Load cell: The load cell associated with this scale has a capacity of
honeybee colonies. Degenfellner and Templ (2024) used hive weight 150 kg, making it suitable for weighing hives, which can weigh
measurements in combination with machine learning tools to assess up to 80 kg.
colony status and bee health. In their study, an analysis and prediction • BR80 display: The scale is equipped with a Baxtran BR80 dis-
of hive weights was performed, providing a warning system that is play (Baxtran, 2024). This display has six digits, allowing it to
enabled when there are significant differences between predictions and display up to 100,000 different values. In addition, the resolution
actual measurements. can be adjusted from 1 kg to 5 g. In this case study, a resolution
Under these considerations, in this work, we use measurements of of 100 g has been selected for the subsequent analysis of the
the weight of a hive and machine learning algorithms to determine the measurements, as this was considered more than sufficient.
different stages of the flowering period. • Transmission interface: The BR80 display is connected via a DB9
connector to an RS-232 interface. The transmission frame consists
3. Characteristics of the measurement system of seven bytes. The first character sent is an equal sign ≪=≫,
which serves to synchronize with the Waspmote. Then, the six
A monitoring system (Fig. 1) designed for an apiary with 6 colonies digits relating to the weight value are sent, which are encoded
is described by Gil-Lebrero et al. (2016). The apiary is located on in ASCII code. The transmission speed of these values can be
a campus of the University of Cordoba (Southern Spain). The area configured between 1200 and 9600 baud, the 1200 baud rate
presents a Mediterranean climate, mostly arid and cloudless, with little being chosen to avoid possible overflows in the reception buffer
rainfall. Specifically, winters are soft and summers are extremely hot. of the Waspmote’s UART, when sending these data together with
The system, named Wbee, allows remote monitoring of the weight, those of the three SHT15 sensors.
temperature and humidity of the hives. The architecture of the system
is shown in Fig. 2 and is divided into three levels. Wireless nodes Near the beehive apiary there is a 230 VAC main power supply.
that monitor bee hives and send data are at the lowest level. The Each beehive system required two power sources. One was a 5 VDC/1 A
intermediate level consists of a computer with an application that supply via miniUSB for the Waspmote, and the other was a 12 VDC/0.5
collects the information sent by the nodes in each hive and stores it A supply with a female jack for the weighing scale. The scale housed a
in a local database. At the top level of the architecture is an-Internet 6 V 2500 mAh Li-ion battery, while the Waspmote used a 3.7 V 2300
data server that collects the information from all the hives and contains mAh Li-ion battery. Both devices featured integrated charging circuits,
a copy of the local database. allowing battery recharging from the external supplies and ensuring
The wireless node is based on the basic version of Libelium’s Wasp- continued operation during power outages.
mote (Libelium, 2024) (Fig. 3(a)). For this purpose, a specific adapter In the apiary, the local computer plays a crucial role in acquiring
board has been designed which is inserted into the Waspmote. The and processing data from the bee hives. This computer runs two appli-
complete system consists of the following components (Fig. 4): cations: SCADA (Supervisory Control And Data Acquisition), which is

3
A. Gersnoviez et al. Computers and Electronics in Agriculture 237 (2025) 110508

Fig. 2. Architecture of the monitoring system.

some sets of daily measurements of specific hives had to be eliminated

because of errors or because they were incomplete. With all this, the
final number of sets of daily measurements in different hives, which
will be used for further processing, is 279.
By analyzing the evolution of the weight of a hive throughout the
flowering period, it is easier to see the different stages of flowering. As
an example, the evolution of the weight of Hive 4 in the 2016 flowering
period can be seen in Fig. 5. It can be seen that, before the beginning of
flowering (pre-flowering stage), the weight during these days remains
practically constant; when the flowering stage begins, the weight starts
to rise and remains so until the end of flowering; from then on, the
weight remains constant again (post-flowering stage).
This information is very useful to know which days belong to
Fig. 3. Detail of the monitoring system: (a) One of the nodes installed in the beehive; the different stages and to analyze the measurements related to each
(b) Weighing scale. of these days. The objective is, with the measurements taken on a
particular day, to determine in which stage of flowering the hives are
in. And, not only the objective of this study is to determine if they
responsible for synchronizing, requesting and processing the data from are in pre-flowering, flowering or post-flowering, but also, in case they
each hive node, and MySQL, a free software used to manage the local are in full flowering, if they are in an initial stage (which we will
database. In addition, there is a global server that also runs MySQL and call flowering stage-1) or a final stage (which we will call flowering
replicates the local database for each hive. This ensures an additional stage-2).
level of security and allows data to be available in the cloud. After labeling the days in the pre-flowering, flowering and post-
Data acquisition is performed periodically. Each hive node remains flowering stages, in order to distinguish between flowering stage-1 and
in standby mode and wakes up according to a predefined pattern to flowering stage-2, the last 7–15 days of flowering have been analyzed.
minimize energy consumption. When it wakes up, the node waits for an As can be seen in Fig. 5, in these last days, the slope of the hive weight
acquisition request sent by the local computer via a broadcast message growth varies. It is these days that have been labeled as flowering
to the whole network. All nodes collect weight data synchronously from stage-2 and the previous days as flowering stage-1.
all hives and send the collected data together with their timestamp. Knowing which stage the different days of the measurements belong
If a node does not receive a collection order, it stores the data on to, the different data obtained are analyzed. It is at this point, after
the microSD within the set period. This ensures that the node collects observing how the weight of the hives evolves throughout the day, that
the data within the set time periods, even in case of communication we find interesting information about the phase of the flowering period
problems with the local computer. in which the hives are in. As a proof of this, Fig. 6 shows different
examples of the evolution of the weight of the hives, belonging to
4. Description and analysis of the dataset obtained the different flowering stages. As can be seen, it is possible to find
differentiating elements from one stage to another only from these
During 2016, measurements were taken on 81 days, from 1 May at figures.
00:00 to 20 July at 23:59. The measurements were taken every 5 min, Within the different variables that can be derived from these figures,
obtaining 287 measurements of each of the parameters on each day. we focus our attention on: ‘‘Weight Diff’’, which is the difference in hive
Between days 63 and 68 there are no measurements, due to a serious weight between the beginning and the end of the day; ‘‘Hour min’’,
power supply failure on the campus where the hives are located. At which is the time at which the minimum value of hive weight is reached
the time of the power failure, the system continued to operate thanks to during the day; ‘‘Hour max’’, which is the time at which the maximum
the batteries. However, as the breakdown took several days to be fixed, value of the hive weight is reached during the day; and then two more
the batteries were depleted, hence the lack of data for those days. In variables ‘‘Dur min’’ and ‘‘Dur max’’ are added, which report how long
addition to this, it must be taken into account that, due to data cleaning, these minimum and maximum values are maintained. The purpose of

4
A. Gersnoviez et al. Computers and Electronics in Agriculture 237 (2025) 110508

Fig. 4. Block diagram of the monitoring system.

5. Experimental results

Once the variables to be used have been determined, the next step
is to design the classifier, as well as test its performance.

5.1. Performance of machine learning algorithms for flowering prediction

Different machine learning algorithms can be used to analyze the

data. The methods used in this paper range from simple methods based
on probabilistic calculations to more advanced methods such as neural
networks or decision trees. The specific algorithms to be analyzed are
described below:

• Naïve Bayes (Webb, 2010): It is a probabilistic method, which is

based on the calculation of Bayes theorem and the independence
between sets of attributes to classify new data into different
classes.
Fig. 5. Weight evolution during the flowering period of Hive 4 from 1 May to 20 July • Support Vector Machines (SVM) (Cristianini and Ricci, 2008): It is
2016. a method based on a two-class binary classifier. For this purpose,
the method focuses on finding a hyperplane that is able to linearly
separate the data into a positive class and a negative class.
these last two variables is to determine whether the reaching of the In the case of multi-class classification problems, each class is
minimum or maximum has a peak shape or a more stationary one. usually compared with all the others to exclude them from the
To corroborate whether these variables can be relevant in a classi- original class and thus form respectively the positive class, and
fier, a histogram is made for each of these variables, as can be seen in the negative class (original class against the rest).
Fig. 7. • Neural Network (Gurney, 1997): It is one of the most widely used
Regarding the histogram of the ‘‘Weight Diff’’ (Fig. 7(b)), it can algorithms capable of solving multi-class classification problems
be seen that most of the flowering days have a positive value (either using several neurons stacked in different layers and levels. There
of the first or second stage), while those of pre-flowering and post- are several types of neural networks depending on the transfer
function used by the neurons and the method of adjustment of
flowering are around zero, so it only differentiates between flowering
the parameters associated with each of them. In this paper, the
and non-flowering.
neural network with backpropagation (Haykin, 1998) will be used
On the other hand, ‘‘Hour min’’ (Fig. 7(c)) distinguishes between
to propagate errors from the output layers of the network to the
the four stages: for pre-flowering it is in the afternoon, although this
input layers of neurons.
is not very clear because of its wide deviation; for flowering stage-1,
• Repeated Incremental Pruning to Produce Error Reduction (Rip-
its mean is around 11:30; for flowering stage-2, around 9:00; and for
per) (Cohen, 1995): It is an iterative method of rule induc-
post-flowering at 7:30. The latter three have a narrower deviation, so tion (Langley and Simon, 1995) based on the principle of divide-
it is a good variable for a classifier. and-conquer. Iterative rule generation starts from the least fre-
Analyzing the histogram of ‘‘Hour max’’ (Fig. 7(c)), it has a com- quent class to the most frequent class in order to obtain rules that
pletely random behavior in the pre-flowering stage, but it is more accu- are able to classify more accurately those instances that belong to
rate to distinguish between flowering (both stages) and post-flowering a minority class.
as its mean occurs between 20:00 and 24:00 during flowering and • Partial decision tree (PART) (Ibarguren et al., 2016): This is an-
between 14:00 and 18:00 for post-flowering. other method based on the principle of divide-and-conquer. In this
In the case of the histograms of ‘‘Dur min’’ and ‘‘Dur max’’ (Fig. 7(e) case the method generates a set of rules that form a tree using
and (f)), there is no information to highlight. This is because they do partial trees. The decision to generate branches is based on the
not act as independent variables, but as a reinforcement of the ‘‘Hour metric called Entropy (Fürnkranz, 2010) which helps to calculate
min’’ and ‘‘Hour max’’ variables respectively. the homogeneity of the data.

5
A. Gersnoviez et al. Computers and Electronics in Agriculture 237 (2025) 110508

Fig. 6. Examples of daily weight evolution, where ‘‘x’’ is the initial weight of the hive on that day and, in the legend, ‘‘H– D–’’ refers to the specific hive and day: (a) pre-flowering
stage; (b) flowering stage-1; (c) flowering stage-2; (d) post-flowering stage.

Fig. 7. Dataset information. (a) Number of instances per class; (b) Histogram of Weight Diff attribute; (c) Histogram of Hour min attribute; (d) Histogram of Hour max attribute;
(e) Histogram of Dur max attribute; (f) Histogram of Dur min attribute.

• C4.5 Decision tree (Quinlan, 2014): It is a decision tree generation • SPAARC (Yates et al., 2019): It is a recent method for gen-
method based on the ID3 algorithm (Quinlan, 1986). The opera- erating decision trees based on the optimization and accelera-
tion of C4.5 is based on the subdivision of the training set space tion of the calculation of decision nodes using split-point sam-
pling and node-attribute subsampling. The algorithm uses the
based on the entropy and information gain (Quinlan, 2014) of
simpleCart (Bhargava et al., 2017) as a base algorithm.
each of the attributes of the database, thus generating branches.
• Random Forest (Hsieh, 2023): It is a decision tree based method All the algorithms briefly described above have been coded in the Java
that generates an ensemble of multiple random trees and classifies programming language and using the weka 3.8.6 (Frank et al., 2016)
data mining and machine learning library.
the instances based on the class that most of the random trees
In order to analyze the performance of the different machine learn-
have chosen. The generation of the set of trees is created by a ing algorithms discussed above and to choose the best one, it is nec-
bagging method (Altman and Krzywinski, 2017) by randomizing essary to obtain metrics that are representative of the performance.
and splitting the training set to generate these trees. There are different metrics (Kulkarni et al., 2020) such as confusion

6
A. Gersnoviez et al. Computers and Electronics in Agriculture 237 (2025) 110508

Table 1 Table 2
Algorithm classification performance. Comparison of algorithms according to taxonomy, interpretability and failures to
Algorithm Pre- Flowering Flowering Post- Weighted classify a specific class.
Flowering Stage-1 Stage-2 Flowering average Algorithm Taxonomy Interpretability Classification failures
F1 -Score F1 -Score F1 -Score F1 -Score F1 -Score
Naïve Bayes probabilities pre-flowering
low
Naïve Bayes 0.6753 0.8367 0.7591 0.8400 0.7949 flowering-stage-1
SVM 0.7912 0.8824 0.8070 0.9362 0.8566
parameters flowering-stage-1
Neural Network 0.7160 0.9137 0.8142 0.8627 0.8566 SVM low
optimization flowering-stage-2
Ripper 0.7416 0.8816 0.7414 0.8571 0.8272
PART 0.8000 0.9085 0.8235 0.9167 0.8738 parameters
Neural Network low flowering-stage-1
C4.5 0.8000 0.9346 0.8908 0.9583 0.9061 optimization
Random Forest 0.8049 0.9161 0.8522 0.9020 0.8836 flowering-stage-1
SPAARC 0.7654 0.9007 0.8387 0.9020 0.8661 Ripper rules high
flowering-stage-2
Highest scores highlighted in boldface. flowering-stage-1
PART rules high
flowering-stage-2
C4.5 decision trees high pre-flowering
matrix, Accuracy, Precision, Recall, F1 –Score,... But the last mentioned PART
multiple
medium
pre-flowering
is one of the most usual to summarize results since it contemplates decision trees flowering-stage-2

the calculation of precision and recall being a very robust metric. The pre-flowering
SPAARC decision trees high
flowering-stage-1
calculation of this metric for each class without the need to know how
to calculate the precision and recall metrics can be seen in Eq. (1):
𝑇𝑃
𝐹1 –𝑆𝑐𝑜𝑟𝑒 = (1)
𝑇𝑃 + 1
⋅ (𝐹 𝑃 + 𝐹 𝑁) It can be seen in the last column of Table 1 that the overall
2
performance of all algorithms is over 80% for almost all algorithms.
Let 𝑇 𝑃 be the number of instances that a classifier predicted correctly This shows that it is possible to perform classification of the different
for the analyzed class (true positives), 𝐹 𝑃 the number of instances that stages of flowering. In particular, it is observed that Naïve Bayes is one
the classifier predicted as the analyzed class, but actually belonged to of the algorithms that obtains less accurate results as it is a very simple
another class (false positives), and 𝐹 𝑁 the number of instances that method based on conditional probability of the data observations. In
the classifier predicted as another class, but actually belonged to the addition, it is observed that for each class in general the algorithms
analyzed class (false negatives). based on decision trees are the most accurate obtaining the highest
In addition to the F1 –Score metric, its weighted average variant, results in the F1 –Score metric. In particular, the C4.5 algorithm stands
which is an excellent indicator that summarizes the behavior of an out from the others achieving the highest values even 4% above the
algorithm considering all its classes and the number of instances in other algorithms in the post-flowering class. It can be observed that
each of them in a single numerical value, will be used to carry out a the only case where C4.5 is slightly worse is in the pre-flowering class
comparison between algorithms. The calculation of this metric is shown where the result obtained for this class is only 0.048 difference with
in Eq. 2: respect to Random Forest which is the best model that classifies this
Weighted 𝑛𝐶𝑙𝑎𝑠𝑠𝑒𝑠
∑ particular class. Furthermore, it can be observed that the pre-flowering
𝑛𝑖
Average = ⋅ F1 –Score𝑖 (2) class has the highest complexity due to the fact that it has the lowest
𝑛𝑡𝑜𝑡𝑎𝑙
F1 –Score 𝑖 value in the F1 –Score metric for all algorithms. This is because the pre-
Let 𝑛𝑖 be the number of instances belonging to class 𝑖, 𝑛𝑡𝑜𝑡𝑎𝑙 the total flowering and flowering stage-1 classes are very similar to each other
number of instances covering all the classes, 𝑛𝐶𝑙𝑎𝑠𝑠𝑒𝑠 the number of as can be seen in the histograms of the variables from Figs. 7(b) to 7(f).
different classes covered in the problem, and F1 –Score𝑖 the calculation Table 2 shows a comparative analysis between algorithms focusing
of the metric F1 –Score using Eq. (1) for class 𝑖. on different sections. The taxonomy refers to the method of opera-
Regarding the comparison between the algorithms, a stratified 10- tion to generate the classification models. On the one hand, we have
fold cross-validation mechanism (Browne, 2000) has been carried out. models based on probabilistic methods (naive Bayes) to classify the
This mechanism is applied to the dataset described in Section 4 and data. Methods based on optimization of function parameters to adjust
allows analyzing the behavior of the machine learning algorithms by these functions to the data and to differentiate the classes to label
partitioning the dataset into training and test sets (each fold) while the data (SVM, Neural networks). Methods based on decision rule
maintaining the ratio of the classes to perform the partitions (see Fig. generation (PART and RIPPER). Finally, methods based on decision
7(a)). Specifically, each algorithm is trained and validated with 10 trees, which, although similar to the previous ones, generate a model
different sets generating 10 different models and allowing, on the one of rules following a tree hierarchy.
hand, to analyze its behavior with data with which the algorithm has In addition, this table presents a comparison of the interpretability
not learned, and on the other hand, a certain independence of the data of the different models based on whether a person can understand
and performance results of the algorithms. Therefore, after acquiring how the model works. It can be seen that the models that are more
279 daily data sets detailed in the previous section and applying the interpretable are those based on rules and decision trees, since they
10-fold cross-validation mechanism, each partition uses 251 data for are the ones that perform a classification depending on whether a value
training and 28 for testing. This process is repeated as many times as of an attribute exceeds a certain threshold. Special mention should be
the number of folds, 10 in total. made of random forest because it has a lower interpretability value
Table 1 shows the results of applying the different algorithms since the classification is formed by multiple classification trees and
described above and the F1 –Score metric for each of the classes and the is therefore more difficult to interpret.
Weighted Average F1 –Score metric summarizing the overall performance Finally, Table 2 shows in which classes the most errors are made
of each algorithm. In addition to the metrics presented in this table, by the algorithms. It should be noted that in this table the errors
additional metrics such as Accuracy, Precision, Recall and confusion have not been weighted by the number of existing patterns, but have
matrices have been calculated for each of the algorithms and can been chosen by total errors (this information can be consulted in the
be consulted by the reader in a public repository (Rodriguez-Lozano, confusion matrices for each method available in the repository). It can
2025). be generally observed that most of the models encounter difficulties

7
A. Gersnoviez et al. Computers and Electronics in Agriculture 237 (2025) 110508

To conclude the discussion, it should be noted that, in addition to

the good results obtained by this algorithm, both the processing of the
data to obtain the values of the classifier input variables, as well as the
operations to go through the tree in Fig. 9, require simple calculation
operations that can be implemented without difficulty in a simple, low-
cost microcontroller. This microcontroller, together with a display, can
be integrated into the monitoring system to alert the beekeeper to the
stage of flowering of the bee hives.

5.3. Study limitations

The results shown in the previous subsections are promising and the
result of the application of a new approach to modernize the existing
beekeeping techniques, although they are not without limitations.
One of the main limitations of the present work is that it is a study
carried out in a specific geographical location where the possibilities of
rainfall and abrupt changes in the climate during the flowering stage
occur very infrequently over the years, unlike what happens in other
Fig. 8. C4.5 Confusion matrix.
areas of Europe and the world. On the other hand, not all bee species
have to behave in the same way as they have behaved in our study.
To solve these problems, it would be necessary to study new data
separating the pre-flowering stages with flowering stages (flowering- sets from different geographical areas and varying weather conditions.
stage 1 or 2) or encounter difficulties in determining the division However, these data cannot be used in raw form, but must have been
between flowering-stage 1 and flowering-stage 2. It should be noted analyzed and labeled by experts to determine what stage of flowering
that both Neural networks and C4.5 make most errors in a single class.
they are in based on the geoclimatic conditions of their acquisition
However, prioritizing the classification results (as observed in Table 1)
sites. This, unfortunately, is a task that is not available in many places
and considering interpretability, C4.5 is the best choice of the analyzed
at present and given the results achieved in the present study may serve
algorithms.
to encourage other researchers to undertake processing of their data.

5.2. C4.5 Decision tree analysis

6. Conclusions
Now that the performance of the machine learning algorithms has
been analyzed, it is interesting to go deeper and analyze the results In this work, a system for monitoring the behavior of bee hives was
of the decision tree C4.5, which is the best algorithm. To perform the used as a starting point and a study was made of the data obtained by
analysis, the confusion matrix of the algorithm has been graphically the system. Specifically, a relationship was sought between the mea-
represented as shown in Fig. 8. This figure also shows the weighted surements taken and the different stages of flowering. After analyzing
averages of the Accuracy, Precision, Recall and F1 –Score values of this the data set, a correlation was found between the evolution of the daily
algorithm, and all of them reach the value of 91%. weight of the hives and the different stages of flowering (pre-flowering,
It can be seen in the confusion matrix that decision tree C4.5 has flowering and post-flowering), even distinguishing, within flowering,
the most errors in the pre-flowering and flowering stage-2 classes, with its initial and final stages.
10 and 7 errors respectively. This effect is due to that both classes, as With the correlation found, it was determined that the variables
observed in Fig. 7, are very overlapped in most attributes, and since related to the daily evolution of weight that are most relevant when
C4.5 is a decision tree that partitions its nodes based on variables, some classifying are: the difference in weight between the start and end of
of the patterns are not completely separable in their corresponding
the day; the time at which the minimum weight value is reached; the
class, generating these failures. In fact, the pre-flowering, flowering
time at which the maximum weight value is reached; and how long
stage-1, and flowering stage-2 classes differ in the hives in days and
they are in this situation of minimum and maximum. Once the variables
it is very difficult for beekeepers to determine their division on the
had been selected, various artificial intelligence algorithms were tested
spot. As can be seen from the results, the percentage of instances that
to perform the classifier, with the C4.5 algorithm providing the best
the C4.5 model fails to classify is very low even in the minority classes
results.
with few instances. Furthermore, in order to fully understand the faults
This leads to the design of a classifier that, from the daily weight
represented in the confusion matrix, and given that the decision tree is
evolution data of the hives, is able to determine with high precision
fully interpretable, the C4.5 tree is graphically represented in Fig. 9.
Note that in this representation there is an example of a pattern and its in what stage of flowering the hives are in (distinguishing between an
attributes in the upper zone of the tree corresponding to the flowering initial and a final stage of flowering), reaching values in the weighted
stage-1 class to show the interpretability and simplicity of the model averages of Accuracy, Precision, Recall and F1 –Score of 91%.
obtained. The above information demonstrates the possibility of differentiat-
It can be seen from the graphical representation of the tree that the ing, by means of artificial intelligence algorithms, between different
number of branches for the post-flowering class, unlike the rest of the periods of bee activity during the flowering season and the periods
classes, has only a single path. This is because, as can be seen in Fig. before and after these flowering periods. Moreover, it is possible to
7(e), the attribute ‘‘Dur min’’ is the discriminant of most of the instances determine when the flowering is ending, before it is completed. This is
within this class, and therefore does not need more branches to reach very interesting, since it would be possible to design applications, based
this class. Unlike this fact, it can be observed that the number of paths on the results obtained from the above methods, that warn beekeepers
for the rest of the classes is much higher, especially for the case of pre- of the upcoming end of the flowering period, thus being able to orga-
flowering and flowering stage-1, where the complexity of classification nize apiary visits and honey harvesting tasks in advance, optimizing
of these classes is reflected. efforts and making beekeeping more profitable and sustainable.

8
A. Gersnoviez et al. Computers and Electronics in Agriculture 237 (2025) 110508

Fig. 9. C4.5 tree model.

CRediT authorship contribution statement scientific purposes, national decree-law 113/2013 (2010-63-EU direc-
tive). This work was in keeping with directive 2010/63/EU of the
Andrés Gersnoviez: Writing – original draft, Visualization, Su- European Parliament and of the Council of 22 September 2010 on the
pervision, Methodology, Conceptualization. Francisco J. Rodriguez- protection of animals used for scientific purposes and Royal Decree
Lozano: Writing – original draft, Visualization, Validation, Software, 53/2013, of 1 February, which establishes the basic rules applicable to
Formal analysis, Data curation. María Brox: Writing – original draft, In- the protection of animals used in experimentation and other scientific
vestigation. José Moreno-Carbonell: Validation, Formal analysis, Data purposes, including teaching.
curation. Manuel Ortiz-Lopez: Writing – review & editing, Resources.
José M. Flores: Writing – review & editing, Supervision, Investigation, Declaration of competing interest
Conceptualization.
The authors declare that they have no known competing finan-
Fundings cial interests or personal relationships that could have appeared to
influence the work reported in this paper.
This research was supported by the Research and Development
Group Industrial Electronics and Instrumentation (IEI) from Universi- Data availability
dad de Córdoba (Spain). This work was supported by the European
Union European Regional Development Fund [FEDER (2014–2020)] Data shared in public github repository.
under Grant RTA2017-00058-C04-01/03.

References
Statement of animal studies
Albert Raj, A., Kavin, Jeeva, T., Kiran, L., 2022. Effective and efficient honey harvest
This study does not experiment directly with the animals, but is alert system for bee farms. J. Phys.: Conf. Ser. 2325 (1), 012040. http://dx.doi.
limited to acquiring data from different non-invasive measurement org/10.1088/1742-6596/2325/1/012040.
Alleri, M., Amoroso, S., Catania, P., Verde, G.L., Orlando, S., Ragusa, E., Sinacori, M.,
sensors in the hives (e.g. weight). Additionally, the ethics review board
Vallone, M., Vella, A., 2023. Recent developments on precision beekeeping: A
of the Universidad de Córdoba considered that this type of project systematic literature review. J. Agric. Food Res. http://dx.doi.org/10.1016/j.jafr.
did not fall under legislation for the protection of animals used for 2023.100726.

9
A. Gersnoviez et al. Computers and Electronics in Agriculture 237 (2025) 110508

Altman, N., Krzywinski, M., 2017. Ensemble methods: bagging and random forests. Hsieh, W.W., 2023. Decision trees, random forests and boosting. In: Introduction
Nature Methods 14 (10), 933–934. http://dx.doi.org/10.1038/nmeth.4438. To Environmental Data Science. Cambridge University Press, pp. 473–493. http:
Arias-Calluari, K., Colin, T., Latty, T., Myerscough, M., Altmann, E.G., 2023. Modelling //dx.doi.org/10.1017/9781107588493.015.
daily weight variation in honey bee hives. Plos Comput. Biology 19 (3), e1010880. Ibarguren, I., Lasarguren, A., Pérez, J.M., Muguerza, J., Gurrutxaga, I., Arbelaitz, O.,
http://dx.doi.org/10.1371/journal.pcbi.1010880. 2016. BFPART: Best-first PART. Inform. Sci. 367–368, 927–952. http://dx.doi.org/
Astuti, P., Hegedűs, B., Oleksa, A., Bagi, Z., Kusza, S., 2024. Buzzing with intelligence: 10.1016/j.ins.2016.07.023.
Current issues in apiculture and the role of artificial intelligence (AI) to tackle it. Imoize, A.L., Odeyemi, S.D., Adebisi, J.A., 2020. Development of a low-cost wireless
Insects 15, http://dx.doi.org/10.3390/insects15060418. bee-hive temperature and sound monitoring system. Indones. J. Electr. Eng.
Bassford, M., Painter, B., 2016. Intelligent bio-environments: Exploring fuzzy logic Informatics (IJEEI) 8 (3), 476–485.
approaches to the honeybee crisis. In: 2016 12th International Conference on Khalifa, S., Elshafiey, E.H., Shetaia, A.A., El-Wahed, A.A., Algethami, A.F.M., Mushar-
Intelligent Environments. IE’2016, IEEE, pp. 202–205. http://dx.doi.org/10.1109/ raf, S., Alajmi, M., Zhao, C., Masry, S., Abdel-Daim, M.M., Halabi, M.F., Kai, G.,
IE.2016.45. Naggar, Y.A., Bishr, M., Diab, M., El-Seedi, H., 2021. Overview of bee pollination
Baxtran, 2024. BR80 Weight tare indicator. URL https://www.baxtran.com/en/ and its economic value for crop production. Insects 12, http://dx.doi.org/10.3390/
indicators/br80.html. (Accessed 2 July 2024). insects12080688.
Bayir, R., Albayrak, A., 2016. The monitoring of nectar flow period of honey Kiromitis, D.I., Bellos, C.V., Stefanou, K.A., Stergios, G.S., Katsantas, T., Kontogian-
bees using wireless sensor networks. Int. J. Distrib. Sens. Networks 12 (11), nis, S., 2022. Bee sound detector: An easy-to-install, low-power, low-cost beehive
1550147716678003. http://dx.doi.org/10.1177/1550147716678003. conditions monitoring system. Electronics 11 (19), http://dx.doi.org/10.3390/
Bayir, R., Albayrak, A., 2018. The determination of the developments of beehives via electronics11193152.
artificial neural networks. Tehni v Cki Vjesn. 25 (2), 553–557. http://dx.doi.org/ Komasilova, O., Komasilovs, V., Kviesis, A., Zacepins, A., 2021. Model for finding the
10.17559/TV-20160419130812. number of honey bee colonies needed for the optimal foraging process in a specific
Bhargava, N., Dayma, S., Kumar, A., Singh, P., 2017. An approach for classification geographical location. PeerJ 9, e12178. http://dx.doi.org/10.7717/peerj.12178.
using simple CART algorithm in WEKA. In: 2017 11th International Conference on Kotovs, D., Zacepins, A., 2023. GIS-based interactive map to improve scheduling
Intelligent Systems and Control. ISCO, pp. 212–216. http://dx.doi.org/10.1109/ beekeeping activities. Agriculture http://dx.doi.org/10.3390/agriculture13030669.
ISCO.2017.7855983. Kulkarni, A., Chong, D., Batarseh, F.A., 2020. 5 - foundations of data imbalance
Browne, M.W., 2000. Cross-validation methods. J. Math. Psych. 44 (1), 108–132. and solutions for a data democracy. In: Batarseh, F.A., Yang, R. (Eds.), Data
http://dx.doi.org/10.1006/jmps.1999.1279. Democracy. Academic Press, pp. 83–106. http://dx.doi.org/10.1016/B978-0-12-
Cejrowski, T., Szymański, J., Sobecki, A., Gil, D., Mora, H., 2019. Low-power wsn 818366-3.00005-8.
system for honey bee monitoring. In: 2019 Twelfth International Conference on Kviesis, A., Komasilovs, V., Komasilova, O., Zacepins, A., 2020a. Application of fuzzy
Mobile Computing and Ubiquitous Network. ICMU’2019, IEEE, pp. 1–6. http: logic for honey bee colony state detection based on temperature data. Biosyst. Eng.
//dx.doi.org/10.23919/ICMU48249.2019.9006659. 193, 90–100. http://dx.doi.org/10.1016/j.biosystemseng.2020.02.010.
Cohen, W.W., 1995. Fast effective rule induction. In: Prieditis, A., Russell, S. (Eds.), Kviesis, A., Zacepins, A., 2016. Application of neural networks for honey bee
Machine Learning Proceedings 1995. Morgan Kaufmann, San Francisco (CA), pp. colony state identification. In: 2016 17th International Carpathian Control Con-
115–123. http://dx.doi.org/10.1016/B978-1-55860-377-6.50023-2. ference. ICCC’2016, IEEE, pp. 413–417. http://dx.doi.org/10.1109/CarpathianCC.
Cristianini, N., Ricci, E., 2008. Support vector machines. In: Kao, M.-Y. (Ed.), Encyclo- 2016.7501133.
pedia of Algorithms. Springer US, Boston, MA, pp. 928–932. http://dx.doi.org/10. Kviesis, A., Zacepins, A., Fiedler, S., Komasilovs, V., Laceklis-Bertmanis, J., 2020b.
1007/978-0-387-30162-4_415. Automated system for bee colony weight monitoring. Agrofor Int. J. 5 (2).
Czekońska, K., Łopuch, S., Miścicki, S., Bańkowski, J., Szabla, K., 2023. Monitoring of Langley, P., Simon, H.A., 1995. Applications of machine learning and rule induction.
hive weight changes in various landscapes. Apidologie 54 (30), http://dx.doi.org/ Commun. ACM 38 (11), 54–64, URL.
10.1007/s13592-023-01012-0. Libelium, 2024. Waspmote Catalogue. URL https://www.libelium.com/iot-products/
Degenfellner, J., Templ, M., 2024. Modeling bee hive dynamics: Assessing colony health waspmote. (Accessed 2 July 2024).
using hive weight and environmental parameters. Comput. Electron. Agric. 218, Liolios, V., Kanelis, D., Rodopoulou, M.-A., Tananaki, C., 2023. A comparative study
108742. http://dx.doi.org/10.1016/j.compag.2024.108742. of methods recording beekeeping flora. Forests 14 (8), http://dx.doi.org/10.3390/
Edwards-Murphy, F., Magno, M., Whelan, P.M., O’Halloran, J., Popovici, E.M., 2016. f14081677.
B+WSN: Smart beehive with preliminary decision tree analysis for agriculture McMinn-Sauder, H.B.G., Colin, T., Gaines Day, H.R., Quinlan, G., Smart, A.,
and honey bee health monitoring. Comput. Electron. Agric. 124, 211–219. http: Meikle, W.G., Johnson, R.M., Sponsler, D.B., 2024. Next-generation colony weight
//dx.doi.org/10.1016/j.compag.2016.04.008. monitoring: a review and prospectus. Apidologie 55 (13), 1–15. http://dx.doi.org/
Eimanifar, A., Pieplow, J.T., Asem, A., Ellis, J., 2020. Genetic diversity and population 10.1007/s13592-023-01050-8.
structure of two subspecies of western honey bees (Apis mellifera L.) in the Meikle, W.G., Adamczyk, J.J., Weiss, M., Gregorc, A., Johnson, D.R., Stewart, S.D.,
Republic of South Africa as revealed by microsatellite genotyping. PeerJ 8, http: Zawislak, J., Carroll, M.J., Lorenz, G.M., 2016. Sublethal effects of imidacloprid on
//dx.doi.org/10.7717/peerj.8280. honey bee colony growth and activity at three sites in the US. Plos One 11 (12),
FAO, 2008. A contribution to the international initiative for the conservation and e0168603. http://dx.doi.org/10.1371/journal.pone.0168603.
sustainable use of pollinators: rapid assessment of pollinators’ status. Meikle, W.G., Holst, N., 2015. Application of continuous monitoring of honeybee
FAO, IZSLT, Apimondia, CAAS, 2021. Good beekeeping practices for sustainable colonies. Apidologie 46, 10–22. http://dx.doi.org/10.1007/s13592-014-0298-x.
apiculture. AO Anim. Prod. Heal. Guid. (25). Meikle, W.G., Holst, N., Colin, T., Weiss, M., Carroll, M.J., McFrederick, Q.S., Bar-
Flores, J.M., Gil-Lebrero, S., Gámiz, V., Rodríguez, M.I., Ortiz, M.A., Quiles, F.J., 2019. ron, A.B., 2018. Using within-day hive weight changes to measure environmental
Effect of the climate change on honey bee colonies in a temperate mediterranean effects on honey bee colonies. Plos One 13 (5), e0197589. http://dx.doi.org/10.
zone assessed through remote hive weight monitoring system in conjunction with 1371/journal.pone.0197589.
exhaustive colonies assessment. Sci. Total Environ. 653, 1111–1119. http://dx.doi. Meikle, W.G., Rector, B.G., Mercadier, G., Holst, N., 2008. Within-day variation in
org/10.1016/j.scitotenv.2018.11.004. continuous hive weight data as a measure of honey bee colony activity. Apidologie
Frank, E., Hall, M.A., Witten, I.H., 2016. The WEKA Workbench, fourth ed. Practical 39 (6), 694–707. http://dx.doi.org/10.1051/apido:2008055.
Machine Learning Tools and Techniques", Morgan Kaufmann, Online Appendix for Ochoa, I.Z., Gutierrez, S., Rodríguez, F., 2019a. Internet of things: Low cost monitoring
"Data Mining, pp. 107–128. beehive system using wireless sensor network. In: 2019 IEEE International Confer-
Fürnkranz, J., 2010. Decision tree. In: Sammut, C., Webb, G.I. (Eds.), Encyclopedia of ence on Engineering Veracruz. ICEV’2019, 1, IEEE, pp. 1–7. http://dx.doi.org/10.
Machine Learning. Springer US, Boston, MA, pp. 263–267. http://dx.doi.org/10. 1109/ICEV.2019.8920622.
1007/978-0-387-30164-8_204. Ochoa, I.Z., Gutierrez, S., Rodríguez, F., 2019b. Internet of Things: Low cost moni-
Genersch, E., 2010. Honey bee pathology: current threats to honey bees and beekeeping. toring BeeHive system using wireless sensor network. In: 2019 IEEE International
Appl. Microbiol. Biotechnol. 87, 87–97. http://dx.doi.org/10.1007/s00253-010- Conference on Engineering Veracruz. ICEV, I, pp. 1–7. http://dx.doi.org/10.1109/
2573-8. ICEV.2019.8920622.
Gil-Lebrero, S., Quiles-Latorre, F.J., Ortiz-López, M., Sánchez-Ruiz, V., Gámiz-López, V., Pahl, M., Zhu, H., Tautz, J., Zhang, S.-W., 2011. Large scale homing in honeybees.
Luna-Rodríguez, J.J., 2016. Honey bee colonies remote monitoring system. Sensors PLoS ONE 6, http://dx.doi.org/10.1371/journal.pone.0019669.
17 (1), 55. http://dx.doi.org/10.3390/s17010055. Prodanović, R., Brkić, I., Soleša, K., Pelić, D.L., Pelić, M., Bursić, V., Tankosić, J.V.,
Gurney, K., 1997. An introduction to neural networks. Taylor & Francis, Inc., USA. 2024. Beekeeping as a tool for sustainable rural development. J. Agron. Technol.
Hadjur, H., Ammar, D., Lefèvre, L., 2022. Toward an intelligent and efficient beehive: Eng. Manag. ( JATEM) http://dx.doi.org/10.55817/ixvm2800.
A survey of precision beekeeping systems and services. Comput. Electron. Agric. Quiles-Latorre, F.J., Ortiz-López, M., Rodriguez-Lozano, F.J., Brox, M., Flores, J.M.,
192, 106604. http://dx.doi.org/10.1016/j.compag.2021.106604. 2024. FPGA-based bee counter system. IEEE Access 12, 30362–30381. http://dx.
Haykin, S., 1998. Neural networks: a comprehensive foundation. Prentice Hall PTR. doi.org/10.1109/ACCESS.2024.3369045.
Henry, E., Adamchuk, V., Stanhope, T., Buddle, C., Rindlaub, N., 2019. Precision Quinlan, J.R., 1986. Induction of decision trees. Mach. Learn. 1 (1), 81–106. http:
apiculture: Development of a wireless sensor network for honeybee hives. Comput. //dx.doi.org/10.1007/bf00116251.
Electron. Agric. 156, 138–144. http://dx.doi.org/10.1016/j.compag.2018.11.001. Quinlan, J.R., 2014. C4.5: Programs for Machine Learning. Elsevier.

10
A. Gersnoviez et al. Computers and Electronics in Agriculture 237 (2025) 110508

Rodriguez-Lozano, F.J., 2025. Flowering-Stage-ML. URL https://github.com/FJ- Yates, D., Islam, M.Z., Gao, J., 2019. SPAARC: A fast decision tree algorithm. In:
Rodriguez-Lozano/HoneyBee-Flowering-Prediction. (Accessed 15 April 2025). Islam, R., Koh, Y.S., Zhao, Y., Warwick, G., Stirling, D., Li, C.-T., Islam, Z. (Eds.),
Sharif, M.Z., Di, N., Liu, F., 2022. Monitoring honeybees (Apis spp.)(hymenoptera: Data Mining. Springer Singapore, Singapore, pp. 43–55.
Apidae) in climate-smart agriculture: A review. Appl. Èntomol. Zoöl. 57 (4), Zacepins, A., Brusbardis, V., Meitalovs, J., Stalidzans, E., 2015. Challenges in the
289–303. http://dx.doi.org/10.1007/s13355-021-00765-3. development of precision beekeeping. Biosyst. Eng. 130, 60–71. http://dx.doi.org/
Terenzi, A., Cecchi, S., Spinsante, S., Orcioni, S., Piazza, F., 2019. Real-time system 10.1016/j.biosystemseng.2014.12.001.
implementation for bee hives weight measurement. In: 2019 IEEE International Zacepins, A., Kviesis, A., Pecka, A., Osadcuks, V., 2017a. Development of inter-
Workshop on Metrology for Agriculture and Forestry (MetroAgriFor’2019). IEEE, net of things concept for precision beekeeping. In: 2017 18th International
pp. 231–236. http://dx.doi.org/10.1109/MetroAgriFor.2019.8909252. Carpathian Control Conference. ICCC’2017, IEEE, pp. 23–27. http://dx.doi.org/10.
Turyagyenda, A., Katumba, A., Akol, R., Nsabagwa, M., Mkiramweni, M.E., 2025. IoT 1109/CarpathianCC.2017.7970365.
and machine learning techniques for precision beekeeping: A review. AI 6 (2), Zacepins, A., Pecka, A., Osadcuks, V., Kviesis, A., Engel, S., 2017b. Solution for
http://dx.doi.org/10.3390/ai6020026. automated bee colony weight monitoring. Agron. Res. 15 (2), 585–593.
Webb, G.I., 2010. Naïve Bayes. In: Sammut, C., Webb, G.I. (Eds.), Encyclopedia of
Machine Learning. Springer US, Boston, MA, pp. 713–714. http://dx.doi.org/10.
1007/978-0-387-30164-8_576.

11