Honeybees play a vital role in sustaining our agricultural economy and maintaining the ecosystem. A healthy and well spread bee population is crucial for better pollination of local crops as well as non-agricultural flora. The decline in global bee population and increased instances of Colony Collapse Disorder (CCD) have drawn attention of researchers all over the world. Recent technological advancements have impacted the bee-keeping industry in numerous ways, and electronic beehive monitoring has significantly improved over the past few years. Monitoring systems have been developed to observe temperature, humidity and acoustics inside the hive, overall weight of the hive and outgoing/incoming bee traffic to gauge the health of beehives. These monitoring systems aided by various wireless communication technologies make it possible for the beekeepers to monitor a large number of hives continuously, simultaneously, from a distance, and only intervene when required.
The most important characteristic of a monitoring system is the set of parameters used for monitoring. Each commercially available solution makes use of its own set of parameters to determine the health of bees. Most of the research carried out in this area focuses on a small set of two to three sensors in each study, rather than examining a bigger set for its collective usefulness. For communication, the monitoring systems rely on either 3G/4G or WiFi networks which are not accessible everywhere, or on satellite communication which can be very expensive. Despite having a high price tag, most of the monitoring systems provide beekeepers with just the raw data from sensors without any analysis on bee health. Proposed systems in the literature have also not been able to make the most of deep learning algorithms, mostly because the data used for training is collected over a short period of time, and from hives with little geographic diversity. Use of such small datasets with limited variations often leads to inconclusive and unreliable results. Beekeepers, in particular from Australia, have not been able to take full advantage of these electronic monitoring systems because of the aforementioned limitations. The vast landscape with no cellular coverage, and the high associated costs of using such monitoring systems are the major challenges faced by the local honeybee industry.
This work addresses the design and development of a beehive monitoring system capable of long range communication with low power consumption. Appropriate sensors for the proposed system are selected after an extensive review of literature. This selection is based on the relevance of sensor with bee health/activity, suitability for long distance transmission over low capacity channels, and optimal use of power. Extraction of appropriate features from sensor data is the key requirement for remote deployment. Different experiments were performed to evaluate various sensors and their features for their importance, and viability for hive deployment. A total of eight sensor systems were deployed in multiple hives, at different locations, and in varying environmental conditions over a 12 month period. During these deployments, Narrow Band Internet of Things (NB-IoT) was thoroughly tested for its communication feasibility from remote sites. Based on the findings, use of NB-IoT is proposed for low cost and reliable communication from remote beehives. The design of this system has also been made available for other researches to use and improve upon.
The aim of sensor deployments in this study is not only to test different sensors and communication for beehive monitoring, but also to build a quality sensor dataset from beehives deployed at different sites. Beehive data collection is a slow process based on the natural activity and life cycle of honeybees. The harsh environment of remote sites, sensor failures, and communication issues make it a very challenging task. A dataset of 2,170 days of beehive sensor data, weather data, and seasonal information has been collected during this study. The resolution of 144 data points per day in this dataset provides a good picture of daily bee activity, and facilitates the use of machine learning in beehive health monitoring. Random forests are used to evaluate the contribution of different sensors in this dataset, as well as of the performance of monitoring system.
Daily hive weight variations are a crucial aspect of hive health and bee activity. Hive weight is affected by multiple complex internal and external factors. Traditionally, an expensive and difficult to deploy weighing scale is used to monitor the hive weight. This is the first work to propose the use of machine learning for beehive weight estimation. Latest machine learning algorithms were tested for their suitability with beehive monitoring and weight estimation, and modified to make most of the information available in beehive sensor data. This work presents two deep learning models for beehive weight estimation, WE-Bee and Apis-Prime. The features for training and testing these models were selected after an in-depth study of bee behaviour, and the impact of environment on bee foraging activity. WE-Bee uses Long Short Term Memory (LSTM) encoders and decoders with temporal attention, whereas Apis-Prime uses self-attention encoders for the same task. These models were tested on sensor systems and hives which were not part of the training set. The promising results validate the good performance of both networks for unseen data. The hives used for the data collection were allowed their natural variations in colony strengths and forager activity, and were moved to sites at a significant distance from each other to collect geographically diverse data. The diversity of the training data played a significant role in the quality of estimations. Use of these machine learning models has the potential to eliminate expensive beehive weighing scales, and reduce the cost of beehive monitoring systems by more than half.
Evaluation of sensors and contribution of features towards a specific task is important for improving and fine-tuning the design of monitoring systems. This work proposes the use of attention weights of self-attention encoders to evaluate sensors and sensor features, as well as to identify the times of day when sensor data carries most information. This enables a significant reduction in the number of features used for estimation. The equally good results of weight estimation with reduced features signify the usefulness of self-attention encoders for feature selection. These findings not only help assess the bee health/activity remotely, but also significantly reduce the monitoring costs. The estimates about hive weight variations using machine learning provide the beekeepers with important information about the hive without using an expensive weighing scale. The promising weight estimates indicate that the proposed system collects important data from the hive, which can also be utilized for a variety of beehive health monitoring tasks.