Session

Technical Session X: Ground Systems

Location

Utah State University, Logan, UT

Abstract

NovaSAR is a commercial S-band Synthetic Aperture Radar (SAR) small satellite, built and operated by SSTL in the UK. One of its primary mission objectives is to carry out maritime surveillance and monitoring for security and defence applications. An investigation was carried out into comparing and contrasting conventional and new methods to perform automated ship detection in NovaSAR images. The outcome of this investigation could show the potential effectiveness of ship detection using spaceborne S-band SAR for Maritime Domain Awareness (MDA).

The conventional approach is to apply a suitable distribution model to characterise sea surface clutter, followed by the implementation of a fixed threshold, Constant False Alarm Rate (CFAR) detection algorithm. In comparison, a RetinaNet-based convolutional neural network (CNN)solution was developed and trained on an open-source C-band dataset in order to determine the validity of applying non-native training data to S-band imagery. The detection performance was then compared with the CFAR technique, finding that for two selected test acquisitions a CNN-based ship detection algorithm was able to outperform a fixed threshold, CFAR-based method in the absence of native training data. CNN ship detection performance was further improved by applying transfer learning to a native S-band NovaSAR image dataset.

Share

COinS
 
Aug 1st, 12:00 AM

A Comparison of Fixed Threshold CFAR and CNN Ship Detection Methods for S-band NovaSAR Images

Utah State University, Logan, UT

NovaSAR is a commercial S-band Synthetic Aperture Radar (SAR) small satellite, built and operated by SSTL in the UK. One of its primary mission objectives is to carry out maritime surveillance and monitoring for security and defence applications. An investigation was carried out into comparing and contrasting conventional and new methods to perform automated ship detection in NovaSAR images. The outcome of this investigation could show the potential effectiveness of ship detection using spaceborne S-band SAR for Maritime Domain Awareness (MDA).

The conventional approach is to apply a suitable distribution model to characterise sea surface clutter, followed by the implementation of a fixed threshold, Constant False Alarm Rate (CFAR) detection algorithm. In comparison, a RetinaNet-based convolutional neural network (CNN)solution was developed and trained on an open-source C-band dataset in order to determine the validity of applying non-native training data to S-band imagery. The detection performance was then compared with the CFAR technique, finding that for two selected test acquisitions a CNN-based ship detection algorithm was able to outperform a fixed threshold, CFAR-based method in the absence of native training data. CNN ship detection performance was further improved by applying transfer learning to a native S-band NovaSAR image dataset.