March8 2023, 11:30 am, EC4-2101A

*** 11:30am Presenter:  Muhammed Patel ***

Title: Detecting whales from aerial imagery

Remote sensing technologies have significantly advanced ecological monitoring, offering unprecedented capabilities for species detection and conservation efforts. However, the manual annotation of extensive aerial surveys remains a daunting task. In collaboration with the Department of Fisheries and Oceans (DFO), we introduce a novel semi-automated deep learning-based whale detection pipeline. This innovative approach aims to enhance the efficiency and scalability of annotating large-scale surveys by addressing challenges such as weak whale signatures, the presence of lookalikes, and heterogeneity. The seminar will focus on the intricacies of training a small object detection model tailored to these challenges, exploring solutions, and discussing aspects of online learning and domain adaptation for ever-changing environmental conditions.

*** 12:00pm Presenter:  Javier Noa Turnes ***

Title: Long-range spatial dependencies for sea ice semantic segmentation.

Studying sea ice presence in the Arctic is pivotal in the sustainable development of society in northern communities. For ice agencies, synthetic aperture radar (SAR) images are the most prevalent data source to generate ice charts that offer a coarse description of ice concentration and prevailing ice types. While such a product aids analysts, it lacks relevant pixel-level information to locate and describe the ice formation. Therefore, there is a significant effort to obtain automatic methods to generate detailed sea ice maps. The nature of SAR images produces significant class variability that leads to spatial non-stationary statistics, posing challenges for sea ice mapping when using current deep learning architectures that ignore large-scale context. Towards enhancing automated sea ice classification, I am presenting a methodology that makes the most of local and global information and gives treatment to boundaries on semantic segmentation of ice and water. Three blocks are utilized, a convolutional neural network (CNN), unsupervised segmentation, and self-attention from transformer networks. Finally, there are two prediction heads: i) pixel-based, on the CNN output, and ii) region-based, on the transformer output. In this manner, the model learns to capture large spatial feature interactions to tackle non-stationary statistics. The main function of region-based head is to promote region-consistent feature maps for pixel-based prediction.