We are testing the use of flying drones and machine learning to map the main coastal habitat types we find along the coast, such as seagrass beds and macroalgae forests (also known as blue forest habitats). Blue forest habitats provide many different ecosystem services, such as carbon uptake and storage, and are important habitats for many marine species.
The main questions we aim to answer are:
Depending on fly height, drone images and map outputs are at a spatial resolution of down to a scale of 1 cm.
For the moment our models perform excellent in separating broad habitat categories, for example between seagrass and macroalgae. But we are hoping also to be able to distinct between different species of macroalgae, and perhaps also be able to assess the health status of, for instance, an eelgrass bed. Another task is to detect kelp as far deep as possible, which is challenging because of lack of light and turbidity in the water.
When planning ground-truthing, the most important environmental gradients in the study area are covered, such as depths and wave exposure.
Transect routes are selected along these gradients to be sure to detect all species in the area. When moving along the transect, GPS positions are taken within patches of different species of macroalgae, seagrass and other habitats, including non-living objects such as sediments and stones. Only species and objects large enough to be visible in the drone images are registered.
A high-precision GNSS positioning antenna (Leica GS18 T GnS) is used for this, pre-programmed to include all probable species and object classes in the study area. At least 10 (preferably many more) observations of each class is needed to train the Machine Learning algorithms properly.
If the water is cold, wader pants or diving suites are used to keep warm. In deeper water, positions must be taken from a boat, however, the position accuracy decreases with depth.
When mapping coastal habitats, we use either a rotor drone (for example a DJI M210) or a fixed-wing drone (such as a DeltaQuad PRO VTOL). A rotor drone has the benefit of being manoeuvrable and can have low speed and even stand still in the air. Whereas a fixed-wing drone typically have longer flight times and can be used for long-distance flights.
Drones used for coastal mapping can be equipped with different types of sensors. These are Red-Green-Blue (RGB) and multispectral (MS) sensors. Typically, RGB sensors (e.g., SONY RX1RM2 42MP) are used to get an overview of the study area, and for annotation, which means categorization of what you see in the drone images into different habitat types. Whereas MS sensors (e.g., MicaSense Altum 5band) are used to train the machine learning algorithms.
We use either the commercial Pix4D software or the open source OpenDroneMap for the post-processing (stitching and georeferencing of orthomosaics) of the images. An important part of the field procedures is the collection of ground truth data, which is done by skilled biologist using a high precision GNSS positioning antennae (Leica GS18 T GnS).
For the moment we are working on a huge amount of data sampled during our campaigns in Larvik, Runde and other places (see Datasets). The SeaBee partner Norwegian Computing Center are continuously improving their machine learning algorithms as more data are coming in from our different missions along the coast (Liu et al. 2022).
A scientific paper is in preparation entitled “Shallow benthic habitat mapping from drones: a comparison of spectral resolution and label structure” (Garrett et al. in prep.), and a protocol for “Field data collection and drone image annotation for coastal habitat classification and mapping” (Kvile et al. in prep.) is also soon to be submitted.
Recently we also published a paper on “Quantifying seaweed and seagrass beach deposits using high-resolution UAV imagery” (Li et al. 2023).
Datasets for this application are published on GeoNode:
For the moment we are aiming to expand our work beyond that of regular habitat mapping, and are trying to estimate the health status and carbon content of eelgrass beds.
Simon Høydal Sætre and Casper Borger – two Master students from NTNU are writing their theses on these topics using data from Ølbergholmen near Larvik, under the supervision of Kasper Hancke (NIVA), Hege Gundersen (NIVA) and Joseph Garrett (NTNU).
These topics are of high interest to managers in Norway, and contributes to national and international policies such as the Global Biodiversity Framework (“Naturavtalen”) and the European Framework Directive (WFD).
Kelp forests are known for their multiple ecosystem services, including high biodiversity and carbon capture and storage. To be able to conserve and protect these important habitats, it is crucial to know where they are. The main policies that the results of this application could be used to support are: