Marine Biodiversity Classification Using Dropout Regularization
Coastal marine ecosystems are highly productive and diverse, but biodiversity of underwater habitats is poorly described due to logistical and financial limitations of diving and submersible operations. Imagery is a promising way to address this challenge, but the complexity of diverse organisms thwarts simple automated analysis. We consider the problem of automated annotation of complex communities of sessile marine invertebrates and macroalgae in order to automate percent coverage estimation. We propose an efficient fusion technique amongst diverse classifiers based on the idea of "dropout" in machine learning. We use dropout technique to weight each classifier implicitly and for each specie we optimize the region of interest (ROI) for highest accuracy. The preliminary results are promising and show 20% increase in average accuracy (over 30 species) when compared with the best base performance of Random Forest classifiers. The data set along with human "ground truth" annotations are available to the public.