Topic-based habitat classification using visual data

It is now common to quasi-automatically generate acoustic bathymetry and optical mosaics from instrumented Autonomous Underwater Vehicles (AUVs). However, further analysis and interpretation of gathered data is needed to address tasks such as habitat characterization and monitoring. This analysis stage is performed by human experts which limits the amount and speed of data processing. While it is unlikely that machines will match humans at fine-scale classification, machines can now perform preliminary, coarser classification to provide timely and relevant feedback to assist human decisions and enable adaptive AUV behavior. This paper presents a preliminary investigation into using a `bag of features' object recognition system for unsupervised clustering of marine habitat imagery. In addition to directly using the high dimensional signature vectors, we also perform clustering based on a low dimensional topic model of the images. We use an AUV transect in the Great Barrier Reef that covers distinct habitat to illustrate the behavior of hierarchical clustering using both representations. Results suggest that both approaches generate clusters of images that are easily recognizable by humans, with significant computational gains to be made by using a topic-based model.