The ectoparasitic mite Varroa destructor is the most serious threat to honey bees in Italy and worldwide and has played a fundamental role in the decline of honey bee colonies all over the Northern Hemisphere in the last decade. Honeybee mortality rates range from 5% to 30% and the Varroa mite represents one of the main causes. This pathology must be continuously contained on levels of infestation compatible with the survival of bee colonies through periodic pharmacological treatments.
Current diagnostic techniques are based on manual visual inspection of characteristic regions in the body of bee or other time-consuming laboratory methods. Late diagnosis of this ectoparasite causes a number of harmful effects either physical, physiological and pathological at the individual and at colony level. The aim of this project is to find valuable solutions to overcome these challenges. The project aims to minimize the detection time of the critical infestation threshold of varroa in bees colonies and to track such infestation over time. This is achieved by designing an automatic detection system of the level of infestation based on visual deep learning algorithms running as a distributed computing system on hives.
The project contains challenging and innovative aspects. The first aspect is a reliable automatic identification method of static and dynamic varroa markers together with the analysis of the correlation between such distinctive features and the level of infestation of a colony. Thanks to the use of hyperspectral and thermal cameras installed inside the hive, the behavior of the bees will be available as new dynamic markers that can improve current state of the art diagnostic methods. Another aspect concerns the design of energy-efficient deep learning algorithms for the detection of the infestation. Since the hives are not connected to any source of electricity, local microcomputers installed in the hives will be powered by solar panels mounted on top of the hives.
The research envisioned in this project is highly innovative as it is intended to contribute to tackle a real and serious problem such as the Varroa infestation in bee colonies. The improvement of monitoring techniques, making them easier and more precise, in addition to a deep understanding of the multiple behavioural adaptations performed by bees against the mite will allow to develop new ways to control this crucial parasitosis.
The results of this project will have practical significance for the application in apiary to establish mite infestation level automatically and remotely aiming to improve the control strategies involving more precise acaricidal applications, avoiding inadequate or approximate management and the risks of development of secondary pathologies causing the colonies decline and reducing the use of chemicals in beekeeping. Therefore the impacts of the proposed research are wide and potentially applied in multidisciplinary fields.
Potential contributions to the state of the art may be the following:
C1. Scientific and technological impact
As Deep Learning requires training on ad hoc labelled data, which are not currently available or partially match the necessities, the project will provide a multi-task dataset: bee segmentation, pose, activity recognition images and videos will be gathered via a large amount of recordings and expert apidology researchers labelling. Proper transfer learning techniques will be investigated to adapt and grant a solid baseline. Moreover, the proposed algorithms will deal with the following challenge: the presence of dense and overlapped bee populations for Segmentation and Pose Estimation; the continuous overlapping and low bee-to-bee differences for Trajectory Tracking; the absence of environmental information for Activity Recognition. Finally, the new challenge of extending the estimation of the infestation level to Varroa-driven behaviours markers focusing on both internal and external beehive analysis and not only to the visual or thermal recognition of the mite as in the state of the art.
Edge computing is a relatively new distributed computing system that still needs sound evidence of its applicability to vertical domains. Under the technological point of view, the developed system will face the relevant technological aspect concerning how to exploit Edge TPU coprocessors (Coral USB Accelerator) that enable high-speed machine learning inferences on System-on-a-Chip computers in an energy-efficient way. We envision to exploit a microservice architecture based on lightweight virtualization technology tailored to resource-limited hosting nodes, with the function-as-a-service paradigm and annotate functions with meta-information concerning energy availability at all computing nodes. To the best of our knowledge, this management software framework is still not available.
C2. Economic impact
The value of bee pollination for agriculture globally is about $215 billion (https://bit.ly/3bu7yaH) and bees are the most important pollinator that helps ensure the conservation of plant communities and biodiversity in natural ecosystems. The economic return of a system that can safeguard the health of bees is therefore very high in itself. As far as Italy is concerned, according to the National Honey Association, the number of beekeepers is 62944 (site visited January 2020). The proposed monitoring system has a low cost of implementation and therefore accessible also to the many beekeepers (68%) who carry out this activity for self-consumption. Wide adoption of an advanced monitoring system will reduce the loss of honey production.
At EU level beekeeping is considered a key part of the agri-food sector, and the EU provides support helping to keep jobs in our rural areas. Starting on 1 August 2019 and running until 31 July 2022, the EU is supporting varroasis one of the 8 specific measures that are eligible for funding (https://bit.ly/3nyjQBj). The activity of this project goes in the direction of supporting these actions.
C3. Possible use of research infrastructures
The project develops a complete infrastructure that detects the level of varroa infestation that can be used as a pilot by many beekeepers in Italy. The proposed monitoring system being automatic and near real-time can complement the current detection system based on manually communicating to the Banca Dati Nazionale dell'Anagrafe Zootecnica significant events. An approach similar to the proposed one can also be applied to other parasites.
The above identified challenges will be tackled through the research program developed for the project (see the previous section). It is worth to notice that the Principal Investigator and her research group has already obtained remarkable scientific results in the area. Obtained results will be submitted to A, A* conferences in the area of computer vision, pattern recognition, image and video processing and edge computing.