Machine learning to classify adult and juvenile trout hybrids

Background/Question/Methods

Human-mediated rainbow trout (Oncorhynchus mykiss) introductions have led to hybridization with native cutthroat trout species in western North America. In Yellowstone National Park, rainbow x Yellowstone cutthroat trout (Oncorhynchus clarkii bouvieri) hybridization contributes to Yellowstone cutthroat population decline. An essential component of species conservation is identification of native versus non-native individuals, however in this system phenotypic identification of hybrids is difficult, especially for juvenile fish. Phenotypic outcomes vary greatly by individual and ancestry, and currently accurate identification requires expensive and time-consuming genetic analysis. A cheaper, faster identification strategy may be found in machine learning. Previous studies have developed models capable of identifying animals in camera trap images with 99-100% accuracy, and we aim to apply similar techniques to trout conservation. Here, we investigate machine learning image classification as a viable option for identifying hybrid and parental trout from pictures captured in the field.

Results/Conclusions

We initially trained and tested our model on 1500 images of genotyped juvenile trout, with a 60-67% accuracy rate. We have increased the number of juvenile images to 2000 and added adult images (600). We have achieved 70% accuracy in preliminary work with the new adult trout images, and are currently working on rerunning the model with the new juvenile images. Adult trout are easier to identify than juveniles, as mature features tend to be more detailed and differentiating, which explains the higher success with a lower sample size. If this project is successful, our model could become a useful tool for trout conservation.