Figure: Drone imaging can help to manage an aquaculture farm, as demonstrated in the CYBELE project.
In September 2021, several partners of the CYBELE project joined the yearly conference of the European Federation of Animal Science (EAAP) in Davos, Switserland. The 5-day conference connects scientists, people from the livestock industry and other stakeholders through a range of interactive presentations, discussions and workshops. On the 4th day of the conference, a special CYBELE-session was organized as part of the Precision Livestock Farming program in order to share insights from the project with the scientific community. The precision livestock farming-related CYBELE pilots and the project coordinator were present to disclose workflows, challenges and preliminary results. Common challenges could be identified in all pilots, such as the difficulty of real-life data (unbalanced data, missing reference data, technical challenges) and the size of the datasets. The availability of HPC in the project led to the opportunity to develop deep learning algorithms on these big datasets, which would not have been possible on a normal PC due to long computing times and intensive memory usage.
The morning of the CYBELE session started off with project coordinator Steven Davy (Waterford Institute of Technology (WIT), Ireland), who gave an introduction to the CYBELE project. He illustrated the need for a more accessible HPC infrastructure in order to process the big data that is gathered in the agricultural sector in a way that is accessible without in-depth HPC or computing knowledge. This was followed by presentations from the different pilots in the project, which demonstrated his claim.
Dan Børge Jensen (University of Copenhagen (UCPH), Denmark) showed a way to estimate the distribution of live weights of pigs around slaughter based on images from a camera in the barn. He used a combination of a pre-trained convolutional network with a secondary regression model to estimate the current weight and weight distribution of the pigs in the pen.
Franziska Hakansson from UCPH followed with a talk about video-based detection of agonistic behaviour (such as tail and ear biting) in pigs using a convolutional neural network (CNN) and recurrent neural network (RNN). The CNN was used to extract features, which in turn were used by the RNN model to classify agonistic events.
Jarissa Maselyne (Flanders Research Institute for Agriculture, Fisheries and Food (ILVO), Belgium) presented the development of an early warning system for individual pig health based on multivariate analysis. Feed, weight or water parameters can be affected and show correlations with observed problems. However, individual pig differences and several real-world challenges make it hard to obtain a one-size-fits-all model. Some models will be further tested, like Long Short-Term Memory RNN and Rocket.
Filippos Mavrepis (University of Piraeus, Greece) presented a workflow to mitigate boar taint problems (an unpleasant smell and taste of meat that can occur in uncastrated male pigs) in the slaughterhouse. The framework is able to deal with the highly imbalanced boar taint data and predicts the chance of boar taint being presence in a pig on the basis of slaughterhouse and barn data. The workflow focused on explainable artificial intelligence to evaluate the contribution and relevance of the different variables for the prediction of boar taint. Different methods were evaluated to balance the data using upsampling of the minority class and downsampling of the majority class.
Ruben Van De Vijver (ILVO) then talked about the development of a system to track pieces of meat on a conveyor belt between when they are taken off of the line and quality measurements later on the cuts, using non-contact industrial cameras and the DeepSort tracking module. This will make it possible to link the quality of the meat back to the pig it came from for further optimization of the pig production chain.
Bert Callens (ILVO) presented the results of two datasets for meat quality where different quality traits (Colour, IMF, pH, drip, etc.) were modelled based on hyperspectral data, evaluating commercial hyperspectral cameras and point spectrometers in different possible usage scenario's.
Figure: example of CYBELE demos for weight estimation on live pigs via cameras and for meat quality estimated on pork via hyperspectral imaging.
Ruben Van De Vijver (ILVO) then presented the results of a classification model which was trained on a dataset of 6600 images collected from over 23 fish species. The deep learning model was further evaluated to determine which details and areas of the pictures were used to classify the fishes. This can be compared to the experts' classification to determine if the model also focuses on the typical characteristics the experts use to identify the fish species (place and shape of fins for example). These models can be the basis for online and real-time classification of fish.
Finally, Steven Davy (WIT) then presented how to use drone imagery, image processing and data mining to optimise the efficient use of feed in aquaculture on coastal farms by modelling the growth and identifying waste. These growth models can then be used to determine the optimal feed level.