Automatic seabed classification using functional data analysis and time series cluster techniques

Abstract

Seabed characterization in coastal environments is usually based on acoustic techniques. Since intrusive measurements are very time-consuming, data acquired by echosounders are the best option for classification purposes. The acoustic seabed response is measured by recording local averages of the intensity field during a time interval, which contains the first echo produced by a sonar pulse excitation emitted from the water surface. The standard methodology for the sea bottom classification relies on the accurate extraction of features, which enable a classical multivariate cluster analysis. The effectivity of such reduction of dimensionality on the data may be enhanced by a preprocessing of the signals based on physical knowledge about the acoustic behaviour of the intensity curves depending in the relative position of the echosounder with respect to the seabed. The automatic seabed classification proposed in this work is performed by means of either time series cluster methods or functional data analysis (FDA) non-hierarchical cluster techniques. In both cases, this method does not require any a priori knowledge of the feature extraction on the sonar curves. More precisely, unsupervised methods such as the FDA K-means method, the multivariate medoids cluster, and time series cluster techniques have been applied. The supervised FDA techniques such as functional generalized linear models (GLM) and generalized spectral additive models (GSAM) have been also considered. The proposed technique is illustrated with some sonar data measured in a controlled environment (where the real classification is well-known) and compared with those results obtained with classical multivariate hierarchical cluster tools.