Attention-Guided Convolution Neural Network Assisted With Handcrafted Features for Ship Classification in Low-Resolution Sentinel-1 SAR Image Data

Abstract

Automatic maritime target classification is essential for effective coastal surveillance and maritime applications. Synthetic aperture radar (SAR) data have been increasingly used for ship classification due to the recent wider open-source data accessibility and the development of deep learning techniques. The convolution neural network (CNN) is a popular deep learning technique used for ship classification using high-resolution labeled SAR image data, despite the limited accessibility of these data due to the substantial cost. The labeled low-resolution SAR image data are less expensive, but their utility is limited due to high false alarms and poor classification accuracy caused by the difficulties in recognizing and discriminating the discrete features for specific classes. To overcome these constraints, we propose an attention-guided CNN model assisted with handcrafted features for ship classification in low-resolution SAR images. This model was trained with OpenSARShip data and tested on Sentinel-1 20 m images with Automatic Identification System (AIS) data class labels from five ports across the world. For the four types of ships (tanker, bulk carrier, container ships and fishing vessels), the experimental results showed an overall classification accuracy of 87.5%, a Kappa score of 0.75, a precision of 86%, and a F1 score of 86.3%. The intercomparison analysis on test data showed statistically a higher performance for the proposed model than for the baseline VGG16 model, which indicates the effectives of the proposed model for various maritime applications.