by Christian Nilsson, University of Gothenburg

Vertical zonation in marine environments occurs when sessile species are primarily found at specific depth intervals, dividing the habitat into distinct horizontal layers. The community assemblage and size of these different zones is determined by environmental conditions as well as interactions between organisms. This phenomenon has primarily been studied in intertidal waters, but also occurs at greater depths, beneath the halocline. Here, low spatiotemporal variation in environmental conditions suggest increased ecosystem sensitivity to recent climate change. However, subtidal environments and their zonation patterns have been understudied historically due to technical limitations. Therefore, in this study we apply machine learning based object detection software to ROV footage to estimate and quantify vertical zonation and community structure along the sides of the Koster submarine canyon. We also plan to investigate possible changes in zonation structure as well as any possible correlations to changing environmental conditions over the past 30 years. For more information - have a look at this conference poster.

by Frida Persson, University of Gothenburg

In this project our student went to South Africa to sample shark observations and population density data, together with the Sharklife Conservation Group. We now have the first models trained to detect up to six shark species in the marine footage recorded with baited remote underwater video systems (BRUVS). The models may be used in the future to discover new and continuously monitor existing elasmobranch hotspots that need special protection from fisheries. The full report is available under Publications. 

by Saga Alsterlind, University of Gothenburg

As the global demand for green energy continues to rise, so does the market for offshore wind farms (OWF). The implementation of the submerged OWF foundations introduces new hard substrates on which marine organisms begin to colonize, also known as the “reef effect”. OWF's therefore have great potential to increase marine biodiversity by elevating habitat value. This project aims to develop a method for quantifying and monitoring changes of sessile biodiversity in relation to OWF's, utilizing machine-learning to perform segmentation detection on still imagery. This project is a collaboration between University of Gothenburg and the renewable energy company OX2.

by Joel Hjärne Kokk, University of Gothenburg

In this project we investigated the efficiency and reliability of the machine learning (ML) approach compared to manual counting of habitat building species. We trained and tested the performance of a object detection model including three species: Desmophyllum pertusum (cold water coral), Bolocera tuediae (deplete sea anemone) and Geodia barretti (football sponge). We then used the model to see whether the distribution of these three species had changed over the last twenty years in one of the reef areas of Kosterhavet National Park. The results highlight the potential of ML models as an analytical tool, but also identify issues that needs to be addressed to make the system applicable for large-scale studies on deep water fauna such as better standardization of subsea imagery. The mapping tool developed during the project is available here.

by Sarah Al-Khateeb and Lisa Bodlak

Large and diverse data is crucial to train object detection models properly and achieve satisfactory prediction performance. However, in some cases, for example when developing models for detecting rare species data will always be limited resulting in overfitting and poor model performance. Furthermore, underwater images can often be of poor quality, with colour distortion and blurriness. In this study we explored methods to both increase and enhance training data for object classes with little training data and thereby improve detection accuracy. We experimented with several Generative Adversarial Networks (GANs) to enhance and increase the training data. The results prepared for publication and the synthetic images are published here.