Dr Martin Skjelvareid1, Dr Katalin Blix1, Dr Eli Rinde2, Dr Kasper Hancke2, Prof Galice Hoarau3
1UiT – the Arctic University of Norway, Tromsø, Norway, 2Norwegian Institute for Water Research (NIVA), Oslo, Norway, 3Nord University, Bodø, Norway
Shallow water habitats are challenging to map. In satellite images of shallow water areas, the resolution is relatively low, and pixels from areas along the coastline contain significant amounts of terrestrial “noise”. Field mapping of these areas (walking, snorkeling, diving, using boats) is also often challenging due to the shallow water and exposure to wind and waves.
We explore the use of unmanned aerial vehicles (UAVs, “drones”) as an alternative mapping tool. The UAVs are relatively light and easy to deploy in the field, and due to the low flying altitude, image resolution is high.
A hyperspectral camera mounted on the UAV captures up to 300 bands in the 400-1000 nm wavelength range, enabling the collection of more detailed spectral information than e.g. standard RGB cameras.
Hyperspectral images and field observations have been collected at several locations along the Norwegian coast, and the images have been annotated with habitat types, including seagrass bed, maerl bed, rockweed, kelp forest and turf algae. In some cases, annotations have also been made of plant species, density, and health condition.
We show that by using machine learning, combined with training data representing a suitable range of variation (location, water depth, weather conditions etc.), it is possible to produce accurate high-resolution habitat maps using our imaging system. By collecting datasets with high resolution both spatially and spectrally, we are also able to simulate how such mapping can be scaled up to systems with lower resolution but higher area coverage.
Biography:
Biographies to come
