Investigation of Object-Based Ship Detection Methods using Range-Compressed Airborne Radar Data

Kurzfassung

Near real-time ship detection and monitoring are essential for ensuring safety and security at sea. Traditional ship monitoring systems are the automatic identification system and marine radars. However, not all ships are equipped with an automatic identification system transponder, and the marine radars suffer from limited visibility. To overcome these shortcomings, airborne radars offer great potential as an additional data source for ship detection in operational use. Existing ship detection methods are hereby mainly applied on fully focused synthetic aperture radar images. The processing effort required for generating fully focused images makes such methods generally unsuitable for near real-time applications. For this reason, within German Aerospace Center’s EMS-HR project, range-compressed airborne radar data are used for ship detection as they do not require conventional time-consuming processing. As the state-of-the-art Constant False Alarm Rate (CFAR)-based algorithms are used for ship detection. Such algorithms have difficulties in achieving near real-time detections due to their pixel-based approach. Besides, since in high-resolution radar data, a ship is typically composed of several pixels, applying pixel-based approaches can result easily in hundreds of detections per ship object, which makes additional post-processing necessary. As CFAR-based algorithms only consider the amplitude value on a pixel level, pixels with similar high amplitude values are falsely detected as ship pixels, while ship pixels with low amplitude values remain undetected. Therefore, object-based ship detection methods using range-compressed airborne radar data are investigated within this master thesis. Initially, investigations were carried out towards Object-Based Image Analysis (OBIA), which considers a group of spatially related pixels as an object, that can be distinguished by its object characteristics from its surrounding. Besides, the Faster Region-based Convolutional Neural Network (Faster R-CNN) was investigated as another object-oriented detector, which learns object characteristics in an automated manner without the need for hand-crafted feature extraction as required by OBIA. Several different Faster R-CNN models were implemented with the well-established ResNet-50 backbone and trained on 21,914 real X-band range-compressed data patches containing at least one ship signal. While the implemented OBIA workflows show several difficulties, especially due to the characteristics of the range-compressed data, the implemented Faster R-CNN models demonstrate high accuracy and robustness, even for very dense multi-target scenarios in the complex inshore environment of the German Bight, North Sea. The Faster R-CNN models reached an overall recall of up to 90.25%. Compared to the CFAR-based ship detector, the Faster R-CNN models indicate fewer false detections and present great potential towards near real-time applications.