Skip to main navigation menu Skip to main content Skip to site footer

Beluga whale detection from sliced aerial remote sensing images using object detection pipelines


The emergence of high-resolution remote sensing imagery greatly facilitates the activities related to conservation biology, including whale counting. As manual annotating is laborious and subjected to human-induced bias, it is necessary to introduce automatic approaches for whale detection from the large remote sensing dataset based on machine learning-based techniques. In this paper, we implement two deep neural network-based object detection models (i.e., RetinaNet and faster RCNN) to detect the presence of whale in aerial remote sensing images obtained from a survey conducted on Cumberland Sound Bay, Nunavut in 2014. To tackle the difficulties in effective detection caused by the sparse occurrence of whales in the large image, an image-slicing approach is adopted to increase the ratio between the size of whale sample bounding boxes and the input image of the model. Testing results show that compared to downsample on the original image directly, the proposed image slicing approach boost the detection accuracy significantly.