In the context of the current global biodiversity crisis, several whales and cetacean species are among the most endangered animals on Earth due to past extensive whaling and to the current impact of human activities. This thesis aims to design an effective and generalizable algorithm to detect, count and localize whales in aerial imagery (drones, planes) with high Ground Sampling Distance (GSD) as well as in satellite images with low GSD, taking advantage of the automation presented in some computer vision architectures. Thus, we have created a large dataset of aerial/drone images in which we have performed a comparison between three computer vision architectures, obtaining a great 0.974 mAP (PASCAL VOC metric) in such type of images. Then, we have transferred the learning obtained in this first stage to detect whales on another dataset with other GSD features composed of only satellite images, reaching a reasonable 0.74 mAP (PASCAL VOC metric) and detecting 74% of the whales. Additionally, we have tested the effect of synthetic images, created ad-hoc from aerial imagery by applying resampling and standard data augmentation techniques. The use of synthetic images has been proved to boost the performance of the architectures in satellite images. These results and architectures bring the possibility of building a detection framework that assesses the whale populations and thus provides important information for their conservation.
We propose to a student or multiple students to work on processing techniques using Deep Learning (Convolutional Neural networks, Generative Adversarial Networks, Semantic Segmentation Networks) to detect and classify marine mammals in photographs and satellite imagery. The computational capacity offered by these new tools will allow the scientific community to better study endangered species and to give an adequate and rapid response to face the current biodiversity crisis.
