The Natural Resources Institute Finland (Luke), in cooperation with Loihde, conducted a pilot project of automatic identification and calculation of white-tailed deer with SAS technology. The algorithm was also used to determine whether the image showed a buck, doe, or a fawn. The idea is that automatic identification facilitates and speeds up reporting and work of researchers.
“At Luke, we monitor game stock, among other things, and make predictions. We monitor different species with different methods and utilize, for example, cameras, follow-up monitoring, and DNA samples. We wanted to test how SAS technology could be used in the identification process and calculation. Examining the images is very mechanical and manual work. When this is automated, it frees up quite a bit of time for the researcher,” says Topi Hanhela, Digital Director at Luke.
The pilot project calculating white-tailed deer began in May 2020, and the system was test-ran a couple of months later in August and September. As part of the same project, Luke and Loihde also tested the development of administrative reporting, as well as the analysis and visualization of commercial fishing statistics using new methods.
The identification and classification of white-tailed deer (buck, doe, fawn) was carried out by an algorithm created in the pilot project with the help of SAS Viya, which was taught identification and itemization step by step. White-tailed deer were selected for the pilot project because a large number of game camera images of them already existed.
“Initially, we only showed images with the white-tailed deer tagged in them to the system. Gradually, through trial and error, the system learned to recognize white-tailed deer in new images. The test data was used to teach the system what went right and what wrong in the identification. The identification rate can be increased quite quickly to over 90 per cent,” says Jens Forsman, Business Advisor at Loihde, who was involved in the project.
Developing an algorithm is quite mechanical. After training material, the system is given test material and the identification rate is increased by training the system. In addition to the white-tailed deer, there are other relevant factors in terms of the system that affect the image, such as the season, different weather conditions, and the amount of light.
“If it is important to identify the desired object at different times of the year and in different lighting, these variations must be introduced to the system’s training material. If, on the other hand, the animal to be surveyed was, for example, a mountain hare, we would have to take into account that the system needs to recognize the animal as a mountain hare, even though it is brown in summer and white in winter.”
Identifications are ultimately mathematical formulas that the system implements. One formula, for example, tells the researcher whether there is a white-tailed deer in the picture, and another one whether it is a buck, and so on.
“The system also tells you the probability of the correct classification. This way a researcher can, for example, check the images where the AI’s identification probability is below a certain percentage.”
Luke produces plenty of data and material as part of its research, statistics, and official work. According to Hanhela, the goal was to create cost-effective and controlled services to enable and facilitate research in line with Luke’s digital strategy.
“When we can automate certain aspects of the research process and, for example, make report preparation easier, we can produce digital services and publications more efficiently for different purposes,” says Hanhela.
Images from game cameras are just one example of potential sources of usable footage. The analysis could be based on a wide range of sources. When it comes to applying the potential of technology, the only limit is imagination.
“For example, overflights and drone images could be one source of footage. The images could be analyzed in the same way as those taken by a game camera. In addition to calculations, image analytics could be used, for example, to observe whether strawberry leaves on a strawberry farm have harmful diseases and how widespread they are,” says Forsman.