In-situ Sea Ice Detection using DeepLabv3 Semantic Segmentation

Identifying type and concentration of sea-ice is a crucial activity performed to assess navigational risk in arctic waters and support daily ice charting activities of the Canadian Ice service. Recent advancements in deep neural networks offer an automated solution to ice charting with retraining capability to continuously improve the system using new data. In this work, we evaluate and compare two popular deep learning based semantic segmentation architectures: Pyramid Parsing Network (PSPNet101) and Deep Labelling for semantic image segmentation (Deeplabv3) for sea ice total concentration detection. The datasets are created using the Nathaniel B. Palmer dataset imagery with four main classes: ice, sky, ocean, and ship, and divided into training, validation, and test sets. The overall mean Intersection of Union (mIoU) performance of Deeplabv3 is 90.21 while for PSPNet101 reported an mIoU of 90.1. The main advantage of the Deeplabv3 model is its compatibility for mobile devices with faster inference speed than the PSPNet101 architecture. The results demonstrate an average inference speed of 0.08s for Deeplabv3 while the PSPNet requires more than 1.9s per image to generate results on a Jetson AGX Xavier deep learning navigation box. Furthermore, the model is capable of successfully performing on mobile devices after Tensorflow Lite conversion with an average execution time of 12s per image on an off the shelf Android device.