HOW TO USE THIS ONLINE MAGAZINE
by clicking the arrows at the side of the page.
by clicking anywhere on the page. A slider will appear, allowing you to adjust your zoom level.
and move the page around when zoomed in by dragging the page.
and return to the original size by clicking on the page again.
by entering text in the search field and click on "In This Issue" or "All Issues" to search the current issue or the archive of back issues
a PDF of this magazine.
by clicking on thumbnails to select pages, and then press the print button.
a page via email, Facebook, Twitter and more.
TO VIEW PREVIOUS EDITIONS
, click the
button at the bottom of the screen.
Beyond the Bale : September 2019
ON FARM 57 ARTIFICIAL INTELLIGENCE POTENTIAL TO AID SELECTION Favourable results have been achieved in an AWI- funded pilot study of whether artificial intelligence (AI) technologies can identify several key phenotypic traits in young sheep. The project concluded that over the next five years the technology has considerable potential to progress to a stage where it could aid selection and productivity management decisions in a commercial wool-growing environment. The aim of the pilot project was to provide an evaluation of whether deep learning AI technologies could accurately predict performance indicator traits of young sheep, from analysing digital photographs of several different views of the sheep. In particular, the project examined whether AI could determine bodyweight, body wrinkle, neck wrinkle, face cover and identify individual sheep. “There has been no other research of this kind done before with sheep,” said lead researcher Dr Mark Ferguson of neXtgen Agri. “It’s exciting that the project provided proof of concept that the technology works. It could potentially be adapted for commercial sheep farming systems and therefore unlock a new horizon for the Australian sheep industry.” “We are confident that in the space of a few years we could develop a model that could outperform humans in both accuracy and speed.” Dr Mark Ferguson, neXtgen Agri The device that was used on-farm by the researchers was a crate that held the sheep, and four fixed cameras – placed above, to the side, in front and to the rear of the sheep – to semi-automatically capture high resolution images and link them to sheep electronic identification (EID). The device also allowed the semi-automatic recording of bodyweight. Using this system, the researchers created an image library of 1,482,041 images from 4,072 sheep. All sheep were weighed at the time of image capture, and subjectively scored for face cover (1-5), neck wrinkle (1-5) and body wrinkle (1-5) and identified to EID. PROJECT RESULTS The AI algorithm could match anonymous face and body images with 94% and 98% accuracy respectively to sheep EID, and 99.7% when both face and body information was used. However, when images from the same sheep were tested five months later, accuracy was considerably lower (<10%) unless images from both time points were included in the training data set (accuracy increased to 90-98%), suggesting that a much larger training data set is needed with repeat images of sheep over time. Using both side and top cameras, the AI algorithms could predict bodyweight with an accuracy of 86% and 87% respectively. Combined information from top and side cameras resulted in an accuracy of 89%. For neck and body wrinkle, the AI algorithms were able to allocate animals to either a high or low wrinkle class with 73%-90% accuracy depending which camera angle and wrinkle trait was predicted. Using the full scale of wrinkle score (1-5) prediction accuracy was lower at 38%-58%. PROSPECTS FOR THE TECHNOLOGY “While AI technology is not yet ready to be deployed in the wool industry, this pilot project has clearly demonstrated that it has considerable potential,” said Professor Raadsma of University of Sydney, who was also a joint lead researcher on the project. “With the correct training data set, machine learning AI models will be very powerful in predicting a range of informative traits from image-based inputs of sheep. We are confident that in the space of a few years we could develop a model that could outperform humans in both accuracy and speed. “As object detection algorithms improve, it could be entirely feasible to collect data in commercial settings from moving objects and make assessments instantaneously allowing for drafting and classification ‘on- the run’. In addition, movement may provide additional information on sheep health and welfare attributes.” “The technology of artificial intelligence lays the foundation for completely new ways to assess traits in sheep without additional time and effort from managers.” Professor Raadsma, University of Sydney The aim of this pilot project was to provide a proof of concept and investment strategy for further consideration. “The benefits from this project will flow in subsequent projects to this one,” said Dr Ferguson. “These benefits include the potential to remotely and automatically weigh and identify animals without extensive infrastructure. It will also lay the foundation for completely new ways to assess traits in sheep without additional time and effort from managers.” Professor Raadsma added: “The concepts initiated in this project will transform the decision-making capacity of the Australian sheep industry in both the tactical management of sheep within a production season as well as the strategic breeding decisions. Once completely developed, the concepts initiated here will augment decisions being made by sheep managers on a daily basis.” Example of head and body capture from the original image taken by the front camera.
In the Shops - September 2019