Data Management Plan

This data package includes datasets and benthic imagery that support the development of a semi-automated coral bleaching classifier in CoralNet. The datasets in this package include annotations of imagery produced to test and train the CoralNet Bleaching Classifier, the annotation label set used in CoralNet, and site-level metadata for annotated imagery. The datasets were generated by scientists of the Ecosystem Sciences Division (ESD) at the Pacific Islands Fisheries Science Center (PIFSC) at NOAA, funded by the Coral Reef Conservation Program (CRCP). This data package includes non-ESD imagery used to develop this classifier, including images of sites in the Papahanoumokuakea National Monument (collected in 2014 by Courtney Couch, a scientist of the ESD), and of sites in Kaneohe Bay, Oahu (in 2015 by The Nature Conservancy). Other source imagery that was used to develop this classifier was collected by the ESD during the 2019 Hawaii coral bleaching event, as well as during the 2015, 2016, and 2019 NWHI National Coral Reef Monitoring Program (NCRMP) survey missions following the traditional photoquadrat collection methodology for NCRMP surveys; this imagery has already been archived, or is a pending archive package; they are described and accessible at https://www.fisheries.noaa.gov/inport/item/59193 and https://www.fisheries.noaa.gov/inport/item/36152.

To create this classifier, named the NOAA ESD Coral Bleaching Classifier (accessible at: https://coralnet.ucsd.edu/source/2947/), a set of training imagery with corresponding point coordinates and label annotations were uploaded to the classifier's source. The training annotations were made by human annotators on both randomly generated and targeted points. Once the training imagery and annotations were uploaded, the automated classification abilities of the classifier were tested by uploading imagery and point locations, but without corresponding label annotations. The machine generated labels for these points were then compared against the human generated labels, which had not been uploaded. See the CoralNet Bleaching Classifier SOP for more information: https://doi.org/10.25923/d0re-9y93.

Main Hawaiian Islands (MHI), including Hawaii, Maui, Oahu, Molokai, and Lanai, and the Northwestern Hawaiian Islands (NWHI), including French Frigate Shoals, Lisianski Island, Pearl & Hermes Reef, Midway Atoll, and Kure Atoll.

Lineage Statement:
To create this classifier, named the NOAA ESD Coral Bleaching Classifier (see 'URLs' in this InPort record for a link to the CoralNet source), a set of training imagery with corresponding point coordinates and label annotations were uploaded to the classifier's source. The training annotations (hard coral, bleached hard coral, and unknown) were made by human annotators on both randomly generated and targeted points. Once the training imagery and annotations were uploaded, the automated classification abilities of the classifier were tested by uploading imagery and point locations, but without corresponding label annotations. The machine generated labels for these points were then compared against the human generated labels, which had not been uploaded. See the CoralNet Bleaching Classifier SOP for more information (link included under 'URLs' in this InPort record).

Process Steps:

• To train a bleaching classifier using CoralNet, the ESD created a new source (CREP-HAWAII BLEACHING v2: Classifier ID 23472) (steps 1-4) containing imagery collected across 178 sites in the Hawaiian Archipelago during the 2019 mass coral bleaching event. A total of 5207 images were used, with approximately 30 images for each site; however, image numbers ranged from 5 - 72 images per site. The imagery was manually annotated using the simple labelset. Once the initial round manual of annotations were completed, a subset of sites with the highest prevalence of coral bleaching were selected and targeted annotations were added manually. CREP-HAWAII BLEACHING v2 served as an experimental test source to confirm that annotations created using steps 5-40 could be uploaded to CoralNet and be used to train a new source.
• To test the machine accuracy of the bleaching classifier, 80% of the sites were designated as training sites and uploaded to the NOAA ESD Coral Bleaching Classifier source using the default feature extractor EfficentNet, while the remaining 20% served as test sites for the source. Sites were randomly sorted. There were a total of 143 training sites selected for NOAA ESD Coral Bleaching Classifier; 129 sites had only random annotations and 14 sites had targeted and random annotations. The resulting 48,622 training annotations were uploaded to NOAA ESD Coral Bleaching Classifier and automated machine classification accuracy was obtained.
• A total of 35 sites were randomly selected as test sites (sites with only random annotations). Test site imagery was uploaded without annotation labels to the NOAA ESM Coral Bleaching Classifier source. Machine suggestions were downloaded and compared to the original human annotations. From this comparison, machine accuracy was calculated as the proportion of machine points that matched with the human annotations.

Quality control is enforced by means of point-to-point, inter-observer calibration exercises that are conducted before each image analysis production series. Training modules and standard operating procedures have also been developed and documented to ensure improved performance and consistent imagery analysis results produced by multiple analysts.

Data can be accessed online via the NOAA National Centers for Environmental Information (NCEI) Ocean Archive.

NOAA IRC and NOAA Fisheries ITS resources and assets.