Monitoring Fish Assemblages in Seasonal Off-Channel Habitats Using Underwater Video and Computer Vision

Abstract

Floodplains provide suitable spawning and rearing habitats for many freshwater fishes and refugia from high flow events. Here, we study seasonal habitat use at the northern edge of the distribution for several spring-spawning fishes in a major Norwegian river drainage and employ underwater video and computer vision to automatically detect, identify, and enumerate species in a seasonal, off-channel backwater slough. Fish actively migrated upriver from their overwintering habitat during the spring runoff, and entered the backwater on the first day it became accessible from the mainstem. A convolutional neural network model was trained to automatically detect species in video obtained via an underwater camera placed at the entrance of the backwater and tested on a representative sample of conditions encountered over the course of a summer season. When we analysed the distribution of prediction scores for tracked fish, we found that the software performed variably for the different species and that the concordance between true counts and software predictions generally improved with increasing mean prediction probability cutoff levels. The intraclass correlation coefficient between the true count and the prediction scores at different cutoff levels showed that the concordance was overall best for roach, followed by pike and tadpoles (frogs and toads). Finally, we found no clear effects of abiotic or optical conditions on the accuracy of the software across a range of prediction probability cut-off levels. We conclude that underwater video provides a feasible, non-invasive means to studying fish in seasonal habitats during vulnerable phases of their life cycle.