23 Oct Image-based Approaches with Artificial Intelligence for Reef Health Monitoring at Nusa Penida Marine Protected Area (MPA)
From the 17th-23rd of October, CTC’s field team conducted their annual reef health monitoring (RHM) at the Nusa Penida Marine Protected Area Learning Site. The RHM combines an underwater visual census (UVC) and benthic monitoring protocols. The UVC records the biomass and abundance of fisheries target species (groupers, snappers etc.) and herbivorous fish (surgeonfish, rabbit fish, parrot fish) that are important indicators of resilience. Currently, CTC uses a method known as the Point Intercept Transect (PIT) method for monitoring benthic reef conditions by recording categories such live hard coral, macroalgae, crustose coralline algae, soft corals, and sponges every 50cm along 3x 50m transects at both 10- and 3-meters depth. The number of times a category is recorded is divided by the number of overall points (in this case 100), which gives us the percent cover of each category per transect. Although this is a longstanding method used by marine scientists, it has limitations because of the skills required by the diver to do the survey, the time limitations from using SCUBA diving equipment, and a lack of statistical power.
This year, CTC trialed a relatively novel approach to their benthic monitoring that combines the use of an image-based underwater photo transect (UPT) methods and automated image analysis using machine learning algorithms known as Deep Learning Neural Networks. The technological development of underwater housings and digital cameras has gradually increased marine scientists’ and conservation practitioners’ capacity to monitor the condition of marine ecosystems like coral reefs using photography. Photographs have been used to estimate the percentage of benthic cover by manual annotating points on an image. The use of this method became popular with the development of coral point count with excel extensions (CPCe) software used to annotate digital images. The problem with using this method previously was the bottlenecks created because of the time to annotate images manually. In Nusa Penida, the team took approximately 3900 50cmx50cm photographic images for recording benthic conditions known as photoquadrats. CTC would need to hire full-time staff to annotate the images and there would also be issues with accuracy because of natural bias and error.
In the last decade, the rise of artificial intelligence (AI) and use of machine learning algorithms similar to ones used for facial recognition have transformed the game for annotating coral reef images. Computer scientists and marine ecologists teamed up to develop Deep Learning Neural Networks that can annotate images up to 700x faster than a human with 80-90% accuracy. This has opened the door for marine conservation organizations that typically lack resources to make use of this technology to annotate their images. CTC will use a recently developed free to use online image repository called Reef Cloud (www.reefcloud.ai). Reef Cloud was developed by the Australian Institute of Marine Science (AIMS) and has built in AI features that automatically annotate points on an image, which quickly give us benthic category estimates. Furthermore, it also allows users to quickly visualize their data to help interpret the data more efficiently.
CTC used the current Nusa Penida RHM to trial this method to understand if we are getting similar results as the PIT method. The photoquadrats were taken on the same transects as the PIT, and so far, the preliminary results suggest the percentage estimates of benthic categories are very similar. This is after expert users have only annotated 5% of the 3900 images that have trained the AI used for the Reef Cloud. The field team is happy with the UPT method and the use of Reef Cloud. We presented the method to the Nusa Penida MPA management unit who were also impressed with this method. This process will allow CTC and other marine conservation practitioners in the region to efficiently monitor their coral reefs and quickly respond to disturbances such as mass coral bleaching events or crown-of-thorns starfish (COTs) (AKA (Acanthaster planci). CTC looks forward to using this method for our RHM in other MPA learning sites such as the Banda and Lease Islands, as well as incorporating other methods that keep us up to date with the latest techniques used to help us monitor and protect tropical marine ecosystems in the Coral Triangle.
Photos: Rosie Poirer & Kasman/CTC