Collection of Shallow Reef Survey Imagery

Lead scientist – Dr Manuel Gonzalez-Rivero

Using diver-propelled vehicles and customised HD cameras (SVII), divers surveyed unprecedented areas of coral reef over multiple reef sites. A depth sensor logger carried by the diver monitored the survey depth every second. Surveys were set to a standard depth of 10 metres (+/- 2 metres) following the contour of the reef at this depth. By standardising the survey depth, we avoided the inclusion of confounding factors (i.e., light, wave energy) when comparing the structure of coral reef communities amongst multiple spatial scales.

During the survey, the SVII camera captured three high-definition images every three seconds during the fifty-minute dive. These images allowed us to continuously record the full 360° environment along a 2-kilometres stretch of reef.

The transect location was selected by looking at capturing a vast variety of coral reef configurations or habitats to best represent the fore-reefs of a region.

read more...

This variability was captured by looking at multiple drivers of coral reef structure, such as environmental regimes (i.e., hurricane incidence, temperature regimes, ocean colour, etc.), local information, geomorphology (atoll, fringing reef, barrier reefs) as well as anthropogenic influence (i.e., coastal development, fishing pressure).

Every image captured for the AXA XL Global Reef Record has a geographical tag that allows us to accurately describe a specific section of reef that can be revisited in time and compared against previously available information. To achieve this, a GPS unit was synchronised in time to the SVII cameras and tethered from the surface by the divers.

In order to accurately and consistently locate the benthic community structure, image scale was required to standardise the area of the reef to be analysed. A customised altitude logger was fitted to the SVII camera, which also synchronised to the time of the SVII cameras. This altimeter, connected to a doppler transducer, logged the altitude of the camera with respect to the seabed and depth. Using this information and the optical properties of the lens of the SVII camera, the footprint area of each downward facing camera was estimated.

Thus, the Global Reef Record provides two main image data-products. The first product is qualitative 360° panoramas of the reef. The panoramas are created by scaling and stitching the fish-eye imagery and provide an interactive way to explore the complete environment of a reef. The second product is a quantitative analysis of benthic categories (% covers) along the reef transects. To generate these covers, from each downward facing image we obtained quadrats with features required for image analysis (colour corrected and cropped quadrats of ~1x 1 m). The image quadrats were re-scaled to fixed pixel / cm ratio and then automatically annotated as detailed next.

Estimating covers of benthic categories

Lead scientists – Dr Oscar Beijbom and Dr Manuel Gonzalez-Rivero

Every field campaign provided about 30-40 thousand survey images which need to be annotated in order to extract benthic cover estimates (e.g., % cover of coral, algae, & sand). Unfortunately, manual annotation of a human expert requires at least 10 minutes per quadrat, which creates a huge bottleneck between collected images and the required data-product (i.e., 30,000 images would take 3 years to annotate). To address this, we developed state-of-the art automated image annotation methods based on deep neural networks.

A random subset of the obtained quadrats for each region was first manually annotated using the random point annotation tool of CoralNet (coralnet.ucsd.edu). This data was then used to train deep neural networks that map image patches to key benthic categories. Label sets of benthic categories have been established based on their functional relevance to coral reef ecosystems and their ability to be reliably identified from images by human annotators. The labels were grouped into main benthic functional components depending on the coral region.
Names and descriptions of labels can be seen here.