Fish Here, Fish There, Fish Detected Everywhere: Developing AI to find Salmonids in Sonar

Salmon and trout (salmonids) are keystone species in the Pacific Northwest, but their populations are threatened by overfishing, habitat loss, and climate change. Accurate population monitoring is essential to understand and address these risks, yet current methods are manual and laborious. Counting by watching sonar imagery is widely trusted, which invites the question – can we teach a machine to interpret this imagery? Sonar presents unique challenges for computer vision: the images are noisy, fish are often obscured or occluded, and environmental conditions change over time. Our research aims to automate detection, tracking, and counting of salmonids in sonar videos using AI. We explore approaches such as self-supervised denoising to improve image clarity, training-free bayesian estimation to distinguish foreground from background, domain adaptation to address differing camera placements, and keypoint-tracking to automatically track areas of interest. Early results are promising, suggesting that automated systems could provide reliable, efficient population estimates and greatly reduce the need for manual counting.