Using Convolutional Neural Networks to Classify Coral Species

Overview

As technology progresses, so does the amount of data collected on the world. Each year the amount of data created continues to grow, both from within the scientific community and from the world at large. With this constant generation, it becomes vital to parse through what has been collected to discover meaningful results. This is especially true of fields such as marine biology. As coral reefs around the world have been in decline over the past couple of decades, it becomes vital to study both the causes of their decline and the corals themselves. The issue arises, however, in how they should be studied. Currently, scientists must look through pictures of corals reefs and classify them with no external help. Classifying just a few hundred photos of corals can take thousands of man-hours to accomplish. In order to help rectify this problem, scientists have begun using convolutional neural networks (CNNs) to help them classify coral species. The methods for how this has been accomplished have varied over the years, but new methods and ideas are constantly progressing the field, and making it easier for scientists to continue their greater goals. A new such method was developed over the course of this research to change the way in which corals are classified. After mimicking and creating various networks for testing purposes, progress began to be made towards the ultimate goal of using CNNs to classify coral species. It began by taking a photo of a section of coral and then splicing that photo into hundreds of smaller photos. The program would then take those smaller snippets and put them into a clustering algorithm which would try to sort them based on similar characteristics that the snippets had with each other. Once every piece had been assigned a label by the algorithm, they would then be reassembled together like the pieces of a jigsaw puzzle. This newly formed picture could then be reused in a different CNN to help it train. Unfortunately, there were issues with the clustering algorithm and the reassembly of the photo. The results of the method are inconclusive as the six-week research project came to a close. This is not to say, however, that the method is useless, but rather that is must be tested further.