Classification of Marine Traffic Activities Using ES Sensors and A VAE-CapsNet Approach

We propose and analyse a new method for the classification of marine traffic activity via deep learning (DL) techniques using capsule neural networks (CapsNets) and variational autoencoders (VAEs). Passive electronic surveillance data is used from 24 hours of activity in Sydney Harbour, Australia, the data is gathered by the electronic support (ES) sensor and extracted to train the VAE-CapsNet model. To address the imbalance in the measured data, data sampling based on oversampling and under-sampling methods is used. According to experimental findings, the proposed system can cover the activity per hour and achieve several classes of 19 with the highest accuracy of 99.19% and less training time than the existing method.