| Summary: | The reliable mapping of the Arctic seabed is a prerequisite for marine spatial planning and environmental management. Underwater imagery (UWI) is one of the most common methods for mapping the seabed. The main advantage of this method is its simplicity, which enables rapid and cost-effective collection of large amounts of data. However, only a small part of information from UWI archives is being used due to labor-intensive and time-consuming effort, often requiring expert knowledge. Thanks to progress in deep learning (DL) methods, with convolutional neural networks demonstrating state-of-the-art in semantic segmentation tasks, new opportunities for automated large-scale analysis of seabed images have emerged. Although some tools exist for intermediate steps, multi-disciplinary teams still lack a unified framework, which could provide fast and reliable semantic segmentation and automate quantitative measurements for classes of interest. To address this issue, we present a prototype segmentation framework tested on the UWI imagery material collected from the upper subtidal zone of central Spitsbergen, Arctic. The system consists of two main layers: a user-friendly web front-end, based on industry leading React framework, and a back-end, relying on TensorFlow Serving API, responsible for managing and serving of DL models. A pilot experiment was conducted using two large 2D mosaics of a seabed transect, stitched from UWI video material, containing 3 classes — pebbles, tube dwelling polychaetes, and Ophiuroidea. Mosaic-based 2-fold cross-validation results revealed a high overlap between ground truth and predicted masks according to similarity coefficients — Dice of 0.86 and 0.84 as well as Jaccard of 0.77 and 0.72 for binary (Ophiuroidea detection) and multiclass segmentation tasks, respectively. Effective, accurate and fast segmentation of the seabed mosaic can be achieved if sufficient training annotations are available for the DL model. Further development is envisaged to enable tweaking capabilities for ...
|
|---|