An object-based seafloor classification tool using recognition of empirical angular backscatter signatures

This study presents a novel concept of seafloor acoustic mapping utilizing the angular dependence of high density soundings. A prerequisite is that data should result from a backscatter-dedicated survey (>100% swath overlap) in order to obtain small-scale seafloor areas ensonified from multiple incidence angles. Accordingly, backscatter data should be geometrically and radiometrically corrected in order to represent only variations due to seafloor type. This method is considered as a mixture of OBIA with empirical ARA and pattern recognition concepts and it provides supervised classification based on empirical backscatter angular signatures of a known set of seafloor types. Therefore it requires a library with all angular signatures corresponding to ground truth locations (seafloor type, dB and angle). The backscatter only needs to be stable and hence this approach is not only applicable on calibrated sonars but works for any MBES system that records backscatter in a stable way. The library should consist of sediment samples, underwater images and/or video which are used to drive the classification and validate its results. Ideally, the ground truth set should cover all different seafloor types from the study area. The concept is that angular backscatter signatures of known seafloor types that have been extracted from fine square areas of seafloor can be utilized for comparison with angular signatures of unknown seafloor. Initially, the study area is segmented into fine squares within which soundings from various beam-angles fall. The smaller the square size, the higher the seafloor homogeneity can be achieved; hence more representative angular backscatter signatures can be extracted for each seafloor type. In this study 5x5 m squares were used for representing naturally homogeneous seafloor. By extracting the angular signatures from the vicinity of sediment sample locations it was possible to use them as reference vectors for performing supervised classification. The classification works in the following way: vectors carrying the mean backscatter value per swath angle are being created from each group of soundings belonging to the same square. Following, each vector is compared to the reference vectors that represent ground-truthed seafloor types. The comparison tests whether the backscatter values of the vector under-comparison fall within a user-defined envelope (range of values) above and below the mean backscatter values of the reference vectors. If the backscatter values for the majority (>85%) of corresponding swath angles belong to the envelope of a reference vector, then these soundings are assigned with the class number of the reference vector. Empirical ARA is more flexible in describing seafloor heterogeneity, compared to physical backscatter models, therefore allowing for classification of a wider variety of seafloor types in a consistent way.