Deep learning-based software for detecting population density of Antarctic birds

Abstract

Monitoring populations of bird species living in Antarctica with current technologies is critical to the future of habitats on the continent. Studies of bird species living in Antarctica are limited due to climate, challenging geographic conditions, and transportation and logistical constraints. The goal of this study is to develop Deep Learning-based software to determine the population densities of Antarctic penguins and endangered albatrosses. Images of penguins and albatrosses obtained from internet sources were labeled using the segmentation technique. For this purpose, 4144 labeled data were trained with five different convolutional neural network architectures TOOD, YOLOv3, YOLOF, Mask R-CNN, and Sparse R-CNN. The performance of the obtained models was measured using the average precision (AP) metric. The experimental results show that the TOOD-ResNet50 model with 0.73 {$AP^{50}$} detects the Antarctic birds adequately compared to the other models. At the end of the study, a software was developed to detect penguins and albatrosses in real time.

Keywords

• deep learning
• antarctic birds
• remote sensing
• population estimation
• convolutional neural network