For our subtidal surveys we used an SLR with a wide-angle lens
mounted on a rigid-frame quadrapod to image shallow benthic and kelp
forest communities at 22 subtidal rocky reef sites (Anacapa Landing,
Arroyo Burro, Arroyo Quemado, Carpinteria, Cathedral Cove,
Cuyler-Bat Rock, Diablo, Ellwood Mesa, Fry’s, Gull Island, Hazards,
Isla Vista, Johnsons Lee-North, Mohawk, Naples, Pelican, Prisoners,
Rhodes, Scorpion, Solimar, West-End and Cat Rock, Yellowbank)
spanning the Santa Barbara Channel and several Channel Islands
National Park monitoring sites at depths equal to or shallower than
60 feet. All surveys were conducted during daylight hours. The
quadrapod image covered a rectangular area of approximately
68cmX45cm.
A team of divers was responsible for collecting the images. At each
site, divers laid out three 20 meter transect tapes. All transects
were marked with GPS start and end points. On all transects one
diver reeled out the transect tape, and a second diver followed the
tape and took pictures on even number meters starting at zero. The
photographer took images of both horizontal and vertical reef
structure on either side of the transect tape (two images per meter
location). In some instances, large blades of brown algae dominated
an image. If this was the case, the diver took a second image of the
same location after clearing the large algae. These images were
labeled as “primary,” meaning nothing was removed, and “algae
removed,” meaning any image obstructing blades were removed before
taking a second image. In some instances, one photo location may
have up to 4 images, two “primary,” and two, “algae removed.”
All images were uploaded directly to BisQue as .CR2 files. Every
image had a grid of 100 circles animated onto it. All 100 circles
were annotated with the centermost species, substrate or no data
category. We focused the annotations on sessile invertebrates and
understory algae cover. Once the images were annotated with 100
points any species not captured in the 100-point analysis were
tagged as present species for occurrence data. Data are organized in
two tables (1) percent cover and (2) occurrence. All annotations
were made manually by technician and undergraduate interns.
From our data set of >1500 images we have obtained >150K data
points with 310 unique tags including individual species, species
groups, and substrates. Identifiable taxa from the following were
annotated, Annelida, Arthropoda, Chlorophyta, Chordata, Cnidaria,
Echinodermata, Ectoprocta, Fish (Gnathostomata), Heterokontophyta,
Mollusca, Phoronida, Porifera, and Rhodophyta. We have obtained very
encouraging results on automated identification of the most abundant
classes (over 80% of data points) with classification accuracy of
over 85%.