Camera traps are widely used to monitor terrestrial fauna. However, platypus pose a particular problem for camera traps due to the small probability of their triggering a camera’s PIR sensor when swimming and the low likelihood of machine learning programs (such as MegaDetector[1]) detecting them in images. Recent advances in platypus monitoring include the use of time-lapse camera traps set to record large numbers of images (normally in the 10's of thousands/site) over a period of days or weeks[2] that then need to be sorted manually to detect platypus[3]. We are developing a machine learning program called Platypus Detector using annotated images of platypus collected from two locations in Tasmania, one lentic (a man-made lake) and three lotic (from a fourth order stream).
We initially trained a YOLO[4] based model with 2,052 images from the four sites to compare AP@0.5 (average precision) of detection. When the model was trained only on the lentic site it performed poorly at the three lotic sites (AP 0.07-0.13). Using a five K-fold cross validation[5] within and between sites results in significant improvements in AP across all sites with AP ranging from 0.48-0.94. By comparison, AP evaluated for images at sites not previously included in training (i.e. training was restricted to the other three sites) ranged from 0.27-0.49. These preliminary results indicate that within-site images are important whilst the model is being developed. However, predicted improvements in AP at unseen sites as additional training sites are added is likely to produce a generally applicable tool for detecting platypus in the future. Further training of the model on a variety of new sites is currently underway.