Transfer learning based ship classification in Sentinel-1 images incorporating scale variant features

Abstract

With the accessibility of Synthetic Aperture Radar (SAR) data and evolution of deep learning, several experiments have been performed in the recent years to actively monitor human activities at sea. Due to the flexible network configuration and complex nature, models based on Convolutional Neural Networks have been able to boost the performance of maritime target classification systems. The novel idea behind these techniques is to enhance the efficiency and automation in the process. However, the limited availability of labeled data and lower resolution of Sentinel-1 (SAR) images have constrained the accuracy of classification models based on the deep learning methods. Besides, it is observed that image processing techniques have resulted in the loss of information present in the images. To overcome these issues, a new method has been developed in the present work that integrates transfer learning techniques (referred to as ResNet50) with additional scale-variant features extracted before image processing. To test our model, SAR and Automatic Identification System (AIS) data were collected from 3 locations: Jawaharlal Nehru (JN) Port (India), Singapore Port (Singapore), and Shanghai-Zhoushan Port (China). A labeling process was done by using the SAR-AIS Neighbourhood integration. Experimental work showed that the proposed architecture of ResNet50 model synergized with extracted scale variant features provides promising results with the classification accuracy of 82.35% and F score of 0.82 on the test dataset, when no models have been developed and trained on Sentinel-1 20 m resolution data for given ship classes.