Leveraging automated image analysis tools to transform our capacity to assess status and trends on coral reefs.

Courtney Couch1,2, Ivor Williams3, Oscar Beijbom4, Thomas Oliver2, Bernardo Vargas-Angel5, Brett Schumacher6 and Rusty Eugene Brainard7, (1)Joint Institute for Marine and Atmospheric Research, Honolulu, HI, United States, (2)NOAA Pacific Islands Fisheries Science Center, Ecosystem Sciences Division, Honolulu, HI, United States, (3)NOAA Pacific Islands Fisheries Science Center, Ecosystem Sciences Division, Honolulu, United States, (4)University of California San Diego, La Jolla, CA, United States, (5)National Oceanic and Atmospheric Administration, Ecosystem Sciences Division, Honolulu, HI, United States, (6)NOAA Pacific Islands Regional Office, Sustainable Fisheries Division, Honolulu, United States, (7)NOAA Fisheries, Pacific Islands Fisheries Science Center, Honolulu, HI, United States
Abstract:
Digital photography is widely used by coral reef monitoring
programs to assess benthic status and trends. In addition to creating a
permanent archive, photographic surveys can be rapidly conducted, which is
important in environments where bottom-time is frequently limiting.
However, substantial effort is required to manually analyze benthic images
which is expensive and leads to lags before data are available. Using
previously analyzed imagery from NOAA’s Pacific Reef Assessment and
Monitoring Program, we assessed the capacity of a trained and widely used
machine-learning image analysis tool – CoralNet coralnet.ucsd.edu – to
generate fully-automated benthic cover estimates for the main Hawaiian
Islands (MHI) and American Samoa. CoralNet was able to generate estimates
of site-level coral cover for both regions that were highly comparable to
those generated by human analysts (Pearson’s r > 0.97 and with bias of 1%
or less). CoralNet was generally effective at estimating cover of common
coral genera (Pearson’s r > 0.92 and with bias of 2% or less in 6 of 7
cases), but performance was mixed for other groups including algal
categories, although generally better for American Samoa than MHI. CoralNet
performance was improved by simplifying the classification scheme from
genus to functional group and by training within habitat types, i.e.,
separately for coral-rich, pavement, boulder, or “other” habitats. The
close match between human-generated and CoralNet-generated estimates of
coral cover pooled to the scale of island and year demonstrates that
CoralNet is capable of generating data suitable for assessing spatial and
temporal patterns. The imagery we used was gathered from sites randomly
located in <30 m hard-bottom at multiple islands and habitat-types per
region, suggesting our results are likely to be widely applicable. As image
acquisition is relatively straightforward, the capacity of fully-automated
image analysis tools to minimize the need for resource intensive human
analysts opens possibilities for enormous increases in the quantity and
consistency of coral reef benthic data that could become available to
researchers and managers.