Automated salmon detection with aerial mapping and deep neural nets
A thorough knowledge about the distribution and numbers of spawning Pacific salmon is fundamental to the conservation and management of its populations. Conventional methods to assess salmon escapement (i.e. the proportion of salmon that returns to freshwater to spawn) can be expensive, invasive and labor intensive.
In this project, funded by the Pacific Salmon Commission, we are working with the Cedar Coast Field station to develop a workflow based on aerial mapping and Machine Learning to automate the enumeration of salmon. In the first stage of this project we focused on Tranquil Creek Chum salmon. More than 4 km of the river was mapped and individual images were stitched together using photogrammetry software. Our first results are promising as salmon are clearly visible in images of stretches of clear water with little surface disturbance.
The resulting orthomosaics were thoroughly scanned and labelled by experienced salmon researchers. We are currently in the process of training Convolutional Neural Nets (CNN) to automate counting using two approaches and different architectures. In the first approach we are using the ResNet-18 architecture and doing scalar regression where the predictor variable is the image and the response variable is the number of salmon in that image. In the second approach we are training YOLOv3 and Faster R-CNN models to detect and localize multiple classes of different categories within an image.
We are grateful that the Pacific Salmon Commissions has granted us further funding to continue data collection and to improve our model so it can be applied to other watersheds with clear water spawning habitat.