Application of artificial neural networks to segmentation and classification of topographic profiles of ridge-flank seafloor

In this paper, we have utilized Artificial Neural Networks (ANN) for seafloor topographic data segmentation and roughness classification using the multibeam-Hydrosweep system (installed onboard ocean research vessel Sagar Kanya) data. Bathymetric profiles from three directions: central (beam 30), port side (beam 10) and starboard side (beam 50), were acquired from the ridge flank and rift valley areas of the Carlsberg Ridge and plain areas of the Central Indian Basin. Self-Organizing Map (SOM) – an ANN architecture employing unsupervised training is used for segmentation of the depth data. This neural architecture is successful in segmenting nonstationary profiles into stationary type before being used for classification. The number of segmented profiles is highest for the rift valley areas whereas for plain area profiles no segmentation is required. Another ANN architecture using supervised training – the Multi-Layer Perceptron (MLP) is applied for the classification of the segmented profiles in two steps. The MLP network was trained using simulated bathymetric profiles of known power spectral (correlation) parameter β, which has been used to classify the multibeam-Hydrosweep depth data. Another parameter S (amplitude parameter) is also being computed for classification. Estimated β and S values of the segmented profiles indicate the sedimentary origin of the plain area. Whereas the estimated parameters for ridge flank and rift values indicate that these areas have volcanic or tectonic origin.