A cluster-classification method for accurate mining of seasonal honey bee patterns

Abstract

Bees are the main pollinators of most wild plant species and insect-pollinated crops and are essential for both plant ecosystems maintenance and humans food production. Among the crops used for human feeding, 75% depend on pollination. In addition to the fact that uncertainty around the beekeeping activity could jeopardize the economic value of pollination, data on honey bee colony losses exist but have not been thoroughly and systematically analyzed to identify potential causal factors. Recognition of seasonal honey bee data patterns can be useful for a number of purposes such as swarming observations, and for forecasting colonies absconding - especially for those hives where the resources are scarce. Here we propose a method to identify honey bee seasonal patterns. The main aim of this research in identifying these patterns is to assist the beekeeper in the management and maintenance of their hives, and, additionally, to prove that with machine learning and, in particular, unsupervised learning is possible to detect seasonal honey bee patterns. We applied a clustering technique in two real datasets from HiveTool.net pursuing brood temperature, relative humidity, and beehives weight. From a clustering validation index and the k-means algorithm, we have found 6 coherent patterns related to seasons. From the found patterns, we compared three well-known classification algorithms (Naive Bayes, k-NN, and Random Forest) to propose a high accuracy classification model (hit rates up to 99.67%) that suggests seasonal honey bee patterns for remote monitoring computing systems.

Introduction

Among all animal pollinators, insects, alone, were valued in €153 billions for their contribution to the pollination of crops worldwide, representing 9.5% of the total value of the world agricultural production used for human food just in 2005 (Gallai et al., 2009). Bees are the most important group of pollinators (Brown et al., 2016) and the Western honey bee (Apis mellifera) is the species most commonly used for pollination purposes around the world (Gallai et al., 2009; Potts et al., 2016). Approximately 75% of crops around the world depend on insects in general for agricultural production of fruits and/or seeds (Ollerton et al., 2011; Potts et al., 2010). Recent works have shown a reduction in the number of pollinating species worldwide (Potts et al., 2016). In particular, honey bee populations have suffered mass deaths in some European regions and in North America due to Colony Collapse Disorder (CCD) and severe winters (Barron, 2015; Gil-Lebrero et al., 2017).

To detect honey bee colonies abnormal states like low adult bee population, spotty brood pattern, and queen loss, it is usually necessary to open the beehives, remove the frames and check on them in a routine called an inspection. In addition to being an invasive process, a careful inspection generates colonies stress (Braga et al., 2020), which can put at risk the pollination services and also honey production. Additionally, bees can be crushed by the frames and box movements. Moreover, many colonies are kept in remote or distant rural apiaries so that inspections at such locations require long shifts. In this sense, remote monitoring of apiaries can assist beekeepers by adding valuable information on the bee's behavior without an invasive inspection (Kridi et al., 2016; Meikle et al., 2017; Murphy et al., 2016; Sánchez et al., 2015; Zacepins et al., 2017; Zogovic et al., 2017), as well as saving the bees from unnecessary stress or other non-productive activities.

Today, thanks to the sensor networks and Internet of Things paradigms, beekeepers and researchers can remotely monitor bee colonies (Kridi et al., 2016; Meikle and Holst, 2015; Zogovic et al., 2017). Remote monitoring via wireless sensors is one of the most important characteristics of the precision beekeeping (Zacepins et al., 2015) which basically involves beehives data collection, data analysis and support decision making in an apiary management context (Dineva and Atanasova, 2018). Once the sensors are installed in the hives, the apiary can be monitored without disturbance, even during periods when invasive inspections of the hives are contraindicated, such as during the winter (Meikle et al., 2017). However, little is yet known about the semantics of the data collected from the hives (Jacobs et al., 2017; Zacepins et al., 2015), such as which physical variables most affect the bees behavior. Such knowledge would help to improve, for instance, the bee colonies' well-being and pollination results.

Here we propose a method for knowledge discovery based on clustering for extracting seasonal honey bee patterns and then get an accurate classification model. Based on this method, our goal is to answer the following central Research Question (RQ): “How to identify biologically relevant seasonal patterns of bee colonies from different hives even if they are from different apiaries?”

The main contribution of this paper is a method that combines clusterization with data classification to detect and recognize seasonal honey bee patterns. To validate the proposed method, we have used two periods found on a yearly basis, the first period corresponds to the spring and summer seasons (the bee active period), and the second corresponds to the autumn and winter seasons in the northern hemisphere, the “quiet” period of the colonies (Kviesis and Zacepins, 2016).

These patterns are composed of temperature, humidity, and weight sensors measurements. To accomplish this, we have established the following activities as showed in Fig. 1:

• obtained raw datasets of temperature, humidity, and weight of the bees' colonies during a full cycle year;

• removed anomalies and normalized data;

• split the colony datasets into subsets corresponding to the seasonal periods of a one-year cycle;

• defined the optimal quantity of seasonal patterns for each period;

• collected the data seasonal patterns for each period;

• recognized and interpreted each data pattern by expert;

• labeled the dataset;

• split the dataset into training, test, and validation subsets;

• applied classification algorithms.