Data-Driven Long Term Change Analysis in Marine Observatory Image Streams

Abstract

In recent years, a number of fixed long-term underwater observatories (FUO) have been deployed to monitor marine habitats over time. HD cameras deployed on FUOs enable vision based studies of long-term processes in the monitored habitats. However, in many marine environments there is often only little a-priori knowledge about potential changes that can be expected or where such changes are likely to occur. Therefore, we propose a method to detect regions of potentially relevant changes and to group them into categories. Wavelet analysis is employed to extract features that describe the approximate progression of pixel values over time. Clustering the features using the recently proposed Bi-Domain Feature Clustering (BDFC) achieves feature grouping and a data-driven definition of change categories. Moreover, a relevance score is computed for each change category, to find regions with relevant changes and to illustrate different relevant change categories simultaneously in one image. Our experiments with images from the Lofoten Vesterålen (LoVe) ocean observatory demonstrate the effectiveness of the method to find relevant change patterns and associate them to different regions or biota.