Classification of inbound and outbound ships using convolutional neural networks

Abstract

In general, a single scalar hydrophone cannot determine the orientation of an underwater acoustic target. However, through a study of sea trial experimental data, the authors found that the sound field interference structures of inbound and outbound ships differ owing to changes in the topography of the shallow continental shelf. Based on this difference, four different convolutional neural networks (CNNs), AlexNet, visual geometry group, residual network (ResNet), and dense convolutional network (DenseNet), are trained to classify inbound and outbound ships using only a single scalar hydrophone. Two datasets, a simulation and a sea trial, are used in the CNNs. Each dataset is divided into a training set and a test set according to the proportion of 40% to 60%. The simulation dataset is generated using underwater acoustic propagation software, with surface ships of different parameters (tonnage, speed, draft) modeled as various acoustic sources. The experimental dataset is obtained using submersible buoys placed near Qingdao Port, including 321 target ships. The ships in the dataset are labeled inbound or outbound using ship automatic identification system data. The results showed that the accuracy of the four CNNs based on the sea trial dataset in judging vessels’ inbound and outbound situations is above 90%, among which the accuracy of DenseNet is as high as 99.2%. This study also explains the physical principle of classifying inbound and outbound ships by analyzing the low-frequency analysis and recording diagram of the broadband noise radiated by the ships. This method can monitor ships entering and leaving ports illegally and with abnormal courses in specific sea areas.