Beluga whale detection in the Cumberland Sound Bay using convolutional neural networks

摘要

The Cumberland Sound Beluga is a threatened population of belugas and the assessmentof the population is done by a manual review of aerial surveys. The time-consuming andlabor-intensive nature of this job motivates the need for a computer automated process tomonitor beluga populations. In this paper, we investigate convolutional neural networks todetect whether a section of an aerial survey image contains a beluga. We use data from the2014 and 2017 aerial surveys of the Cumberland Sound, conducted by the Fisheries andOceans Canada to simulate two scenarios: (1) when one annotates part of a survey anduses it to train a pipeline to annotate the remainder and (2) when one uses annotationsfrom a survey to train a pipeline to annotate another survey from another time period. Weexperimented with a number of different architectures and found that an ensemble of 10CNN models that leverage Squeeze-Excitation and Residual blocks performed best. We evaluatedscenarios (1) and (2) by training on the 2014 and 2017 surveys, respectively. In bothscenarios, the performance on (1) is higher than (2) due to the uncontrolled variables in thescenes, such as weather and surface conditions.