Automatic Classification of Seafloor Image Data by Geospatial Texture Descriptors

Abstract

A novel approach for automatic context-sensitive classification of spatially distributed image data is introduced. The proposed method targets applications of seafloor habitat mapping but is generally not limited to this domain or use case. Spatial context information is incorporated in a two-stage classification process, where in the second step a new descriptor for patterns of feature class occurrence according to a generically defined classification scheme is applied. The method is based on supervised machine learning, where numerous state-of-the-art approaches are applicable. The descriptor computation originates from texture analysis in digital image processing. Patterns of feature class occurrence are perceived as a texture-like phenomenon and the descriptors are therefore denoted by Geospatial Texture Descriptors. The proposed method was extensively validated based on a set of more than 4000 georeferenced video mosaics acquired at the Haakon Mosby Mud Volcano north-west of Norway recorded during cruise ARK XIX3b of the German research vessel Polarstern. The underlying classification scheme was derived from a scheme developed for manual annotation of the same dataset applied in the course of Jerosch [2006]. Features of interest are related to methane discharge at mud volcanoes, which are considered a significant source of methane emission. In the experimental evaluation, based on the prepared training and test data, a major improvement of the classification precision compared to local classification as well as classification based on the raw data from the local spatial context was achieved by the application of the proposed method. The classification precision was particularly improved for rarely occurring classes. In a further comparison with annotated data available from Jerosch [2006] the regional setting of the investigation area obtained by the application of the proposed method was found almost equivalent to the results of an experienced scientist.