Deep learning model for seabed sediment classification based on fuzzy ranking feature optimization

Abstract Accurate acquisition of information on seabed sediment distributions plays an important role in the construction of basic marine geographic databases. Although a multibeam echo-sounding system (MBES) can satisfy large-scale seafloor mapping with high precision and high resolution, the development of a consistent, stable, repeatable and validated seabed sediment classification method based on swath acoustic data is still in its infancy. To achieve accurate prediction and mapping of geographic seabed sediment information, this paper developed a deep learning model based on feature optimization. First, faced with high-dimensional features extracted from multibeam bathymetry and backscatter intensity measurement data, a fuzzy ranking (FR) feature optimization method was proposed. By combining the physical properties of actual sediment samples, the multidimensional features derived from terrain and intensity data are ranked and optimally selected according to the mean square error to eliminate redundant and irrelevant features. Second, the deep belief network (DBN) deep learning method was used to build a supervised seabed sediment classification model. The optimized features and actual sediment samples participate in model training, which further enhances the prediction ability of acoustic data to seabed sediments. Finally, to evaluate the performance of the DBN model, this experiment used large-scale multibeam survey data and ground-truth data (acquired by grabbers, core samplers, dredges, etc.) in the southern Irish Sea to achieve accurate prediction of 10 sediment types (slightly gravelly muddy sand, slightly gravelly sand, gravelly mud, gravelly muddy sand, gravelly sand, muddy sand, muddy sandy gravel, sand, sandy gravel and sandy mud). The experiment results show that by using the optimal feature combination based on FR, the overall classification accuracy and Kappa coefficient reached 86.20% and 0.834, respectively, which are significantly improved compared to the evaluation metrics of other feature selection methods. In addition, compared with the current five typical supervised classification methods (i.e., the random forests, BP neural network, support vector machine, maximum likelihood and decision trees methods), the proposed DBN classification model achieves a better performance, highlighting its application potential in seabed sediment detection and mapping.