Mapping of Algae and Seagrass using Spectral Imaging and Machine Learning

The MASSIMAL project aims to develop new methods for mapping marine underwater vegetation such as seagrass, macroalgae, and maerl. These vegetation types are part of the "blue forests" which are home to a large range of marine species. They also contribute significantly to primary production, carbon capture, and absorption of dissolved nutrients. The blue forests are threatened by human activity, climate change and overgrazing by sea urchins, and new tools are needed for monitoring and studying how and why these ecosystems change. Using a hyperspectral camera mounted on a drone, the seafloor is imaged from 20-100 meters above the sea surface. By combining the hyperspectral images with manual sampling of the vegetation, machine learning algorithms can produce detailed maps of e.g. the different species distribution, vegetation density and physiological state. Data has been collected from numerous locations along the Norwegian coast, in areas close to Bodø, Larvik and Vega. The datasets represent a large variation in plant species, nature types, weather conditions and optical water properties. Hopefully, this variation will enable training of machine learning algorithms that are robust and perform well in many different settings. Several different methods for documenting underwater vegetation and nature types (so-called "ground truth") have been tested during field campaigns. During early stages of the project, ground truth was collected by detailed documentation and imaging of small areas. In the later stages of the project, the methodologies for ground truth collection have shifted towards covering larger areas, and having good examples that span all the variation within an area. This has mainly been done through underwater imaging from a boat (regular or autonomous), combined with logging of precise geolocation, resulting in large sets of geotagged underwater images. The data collected during campaigns must be organized, post-processed and annotated to prepare it for training machine learning models. Developing methodology to do this efficiently and accurately has been an important and time-consuming part of the project. Several data sources need to be combined in order to annotate the hyperspectral images. Annotated datasets will be published as part of the project results. Preliminary results from the project indicate that it's possible to detect distinct spectral and textural patterns for different vegetation and nature types. An early dataset from Bodø has been used to develop methods for feature selection, e.g. selection of the wavelengths that are best suited for vegetation mapping. A master thesis based on data from the Larvik area has promising results regarding mapping of seagrass, turf algae, sand and rockweed. Attempts to use "clustering" on a dataset from Bodø with tightly interwoven classes has demonstrated that semi-supervised learning can be useful in cases where exact annotation is not practically feasible. A "U-Net" machine learning model trained on a dataset from Vega has shown that it's possible to map maerl beds, kelp, seaweed and other algae with relatively high accuracy. The last research campaigns in the project were conducted in 2023, in areas close to Smøla, Larvik and Bodø. Data from the Smøla area showed several examples of how commercial trawling for kelp results in "trawl tracks" across the bottom. Recently trawled tracks are visible as brightly colored stripes with bare rock, while older tracks are dominated by small and young kelp plants. The 2023 Larvik campaign was conducted shortly after the "Hans" extreme weather, and the ocean water had a distinctly brown color and poor visibility due to land runoff. The visibility improved towards the end of the campaign, allowing the collection of good data from areas with seagrass, rockweed and kelp. The 2023 Bodø campaign was a revisit of an area with dense seagrass. As of November 2023, the data from this campaign has not yet been processed. The last two years of the project (2024) will mainly be focused on development of machine learning algorithms and publication of results. A master's thesis published in 2022 has been developed into a manuscript for a journal article, which will be submitted for publication in around the beginning of 2024. A conference article presented at IGARSS 2023 will probably also be developed into a journal article. The large datasets collected during the project will hopefully also form the basis for additional publications. Public outreach will also be given high priority in 2024.

Seagrass meadows and kelp forests are two of the most important marine habitats along the Norwegian coast. These are exposed to stressors such as eutrophication, ocean warming and ocean darkening, which all impact their distribution and health. At present, mapping of these species in Norway is done at a small number of sampling points using underwater “drop cameras”, recording coverage or state parameters of the species at points or along line transects. There is a need for cost efficient tools to map and monitor the distribution and ecological state of blue forests over larger areas and extended time periods. Large-scale mapping based on imaging from satellites or airplanes is possible, but has several drawbacks: Satellites have limited spatial resolution and depend on cloudless days, and airplane missions are costly. We propose using unmanned aerial vehicles (UAVs) equipped with hyperspectral cameras for mapping medium-sized areas. UAVs enable flexible, low-cost imaging missions with high spatial resolution, and hyperspectral imaging will provides detailed spectral information within each pixel. The spectrum of light reflected from underwater vegetation and the seafloor can be used as a “spectral fingerprint” to estimate parameters such as plant coverage, species, biomass and physiological state. Scuba divers, drop cameras and ROVs will be used to aquire “ground truth” measurement of these parameters, and machine learning methods will be used to train mathematical models relating the hyperspectral data to the field measurements. This enables estimation of biophysical parameters for each image pixel. Through statistical analysis of the mapped spatial and temporal changes, we will identify the main drivers that cause the observed patterns. Understanding how the structure and function of these species varies across environmental gradients is essential knowledge for sustainable coastal management.