Recognition of wildlife using deep learning in images taken by camera traps

• Published in: Honyurui Kagaku (Mammalian Science) Vol. 59; no. 1; pp. 49 - 60
• Main Authors: Ando, Masaki, Nakatsuka, Shunsuke, Aizawa, Hiroaki, Nakamori, Satsuki, Ikeda, Takashi, Moribe, Junji, Terada, Kazunori, Kato, Kunihito
• Format: Journal Article
• Language: English; Japanese
• Published: Kyoto The Mammal Society of Japan 2019; Japan Science and Technology Agency
• Subjects: animal recognition; Animal species; Animals; Bears; camera trap; Cameras; Capricornis crispus; Cost analysis; deep convolutional neural network; Deep learning; Image classification; Labor; Machine learning; Object recognition; Researchers; Species; Sus scrofa; Traps; Wildlife; wildlife monitoring
• ISSN: 0385-437X; 1881-526X
• DOI: 10.11238/mammalianscience.59.49

Recently in Japan, camera traps have been used for wildlife monitoring. Modern camera traps operate at a high-performance, thus, over thousands of images can be obtained. Ultimately, this can cause a disadvantage, as checking a large quantity of images is extremely laborious for researchers. This study attempted to decrease the labor involved with the use of the latest technology of image classification, “deep learning”, to recognize the presence of animals, animal species, and the number of heads automatically. Over 110,000 annotated images taken by camera traps were used for constructing our model. In this study, the sika deer (Cervus nippon), wild boar (Sus scrofa), Japanese serow (Capricornis crispus), and Asian black bear (Ursus thibetanus) were targeted. For animal recognition, our model achieved a 15.7% false-positive rate whilst maintaining a 99% true-positive rate. This drastically reduces the quantity of images that researchers must scan through to 43.3% of the original value. When an image of one of the four target species was input, our model successfully returned each species 79.6%, 76.4%, 82.1%, and 76.6% of the time as top-hit category, respectively. For the animal count, when images containing animals were input, each target species reached an accuracy of 91.9%, 84.4%, 91.6%, and 86.4%, respectively. These results suggest that deep learning in camera trap analysis can be a useful tool for reducing the labor cost.