DNN-based seabed classification using differently weighted MBES multifeatures

Seabed sediment classification has significance for the utilization of marine resources and marine scientific research. Currently, the multibeam echo sounder (MBES) is increasingly becoming the tool of choice for large-scale seabed sediment classification. To further explore the technology of seabed sediment classification, this paper proposes a new classification method. In addition to backscatter mosaic, the method also integrates three other different types of features, including texture features of backscatter mosaic, MBES bathymetry features, and backscatter angular response (AR) features, which are given different weights in the classification process. First, geographically weighted regression (GWR) analysis is performed between different types of features and seabed sediment types, and the normalized coefficient of determination (R2) is employed as the weight coefficient for the different types of features. Second, the backscatter mosaic is combined with features from different types to predict the seabed sediment types using a deep neural network (DNN) classifier. Third, the classification residuals of the features from these three different types are acquired through the above classification results. Last, the classification residuals of features from different types are added to the classification results of the backscatter mosaic according to the weights, thereby achieving seabed sediment classification based on MBES multifeatures with different weights. The results show that the overall classification accuracy of the seabed sediments can be significantly improved from 88.98%/85.14% to 93.43% when using the DNN classification model based on MBES multifeatures with different weights compared with the other two models (DNN classification model based on MBES multifeatures with equal weights and DNN classification model based on principal component analysis (PCA) dimensionality reduction). The kappa coefficient can also be significantly improved from approximately 0.85/0.80 to 0.91. Via analysis, the proposed method can reasonably assign the weights of the different features and take advantage of integrating MBES multifeatures for seabed sediment classification. This approach also provides an important reference for future research on seabed sediment classification. •Integrating three different types of feature is more effective for seabed sediment classification.•The GWR model effectively evaluate the importance of the MBES multifeatures.•The DNN-b