Poriferal Vision: Classifying Benthic Sponge Spicules to Assess Historical Impacts of Marine Climate Change

摘要

Sponges and corals are ecologically important members of the marine community. Climate change, while harmful to corals, has historically been favorable to sponges. Sponge population dynamics are studied by analyzing core samples of marine sediment. To date this analysis has been performed by microscopic visual inspection of core cross sections to distinguish spicules (the rigid silica components of sponge skeletons) from the residue of other silica-using organisms. Since this analysis is both slow and error prone, complete analysis of multiple cross sections is impossible. FlowCam® technology can produce tens of thousands of microphotographs of individual core sample particles in a few minutes. Individual photos must then be classified in silico. We have developed a Deep Learning classifier, called Poriferal Vision, that distinguishes sponge spicules from non-spicule particles. Small training sets were enhanced using image augmentation to achieve accuracy of at least 95%. A Support Vector Machine trained on the same data achieved accuracy of at most 86%. Our results demonstrate the efficacy of Deep Learning for analyzing core samples, and show that our classifier will be an effective tool for large-scale analysis.