Title

Machine Learning for the Mapping of Marine Geology, Geomorphology and Habitats

​

Organisers

Riccardo Arosio (University College Cork), Benjamin Misiuk (Dalhousie University), Alexandre Schimel (Geological Survey of Norway)

​

Overview

While marine environments are under increased pressure due to activities such as resource extraction, fisheries, and tourism, there are also calls for increased conservation and more equitable spatial management. Given such competition for the use of marine space, many maritime nations now recognize that high-resolution sea floor data are needed for sustainable management. Developments in acoustic technologies over the past few decades have enabled large-scale surveying of the ocean floor at high spatial resolutions (e.g., metres). Combined with high-quality ground-truth, these data provide the information that is necessary for mapping marine habitats, submarine geomorphology, and geology, enabling effective and sustainable marine spatial management. Alongside developments in marine mapping technologies, machine learning has revolutionized the ways in which environmental data are collected, processed, and operationalized. While uptake of machine learning methods has been rapid for some applications, such as benthic habitat, surficial substrate, and species distribution modelling, development in other related fields has been much slower. Marine geomorphology and geology mapping, for example, is still largely conducted using manual or semi-automated methods. In nearly all cases, the uptake of deep learning techniques, which have potential to increase the automation, accuracy, and objectivity of current mapping workflows, has been slow. There is a strong need for marine geoscientists to network with computer scientists, statisticians, and geomorphologists from other domains in order to foster the development of deep learning methods and workflows.

​

The expansive field of seabed mapping faces several challenges that could be alleviated through the application of novel machine learning techniques. Historically, the work is highly labour-intensive and subjective – depending to a large extent on the training and experience of the data analyst. This may lead to inconsistencies and non-repeatability. Furthermore, the influx of large volumes of high-resolution remote sensing data (i.e., “Big Data”) exceeds the capacity of what may be efficiently processed by human operators. Machine learning, and deep learning in particular, have the potential to greatly increase the efficiency of many seabed mapping workflows including data processing and interpretation, feature engineering, and prediction.

​

With this session at the 4th International Symposium on Machine Learning and Big Data in Geoscience, we aim to foster collaboration between groups working on similar Machine Learning solutions for ongoing seabed mapping challenges, and importantly, also with experts from different but related fields. This may facilitate the transfer of promising methods to seabed mapping. Moreover, we would like to establish an outlet to discuss existing and emerging machine learning approaches and workflows in our field for geospatial modelling, image (acoustic and underwater video/photo) classification, semantic segmentation, supervised, unsupervised, and weakly-, semi- and self-supervised learning. We believe that the International Symposium on Machine Learning and Big Data in Geoscience will provide an ideal platform to address outstanding challenges to the implementation of machine and deep learning for seabed mapping, and to foster valuable interdisciplinary collaborations.

​

​

​