Automatic ship detection based on satellite SAR

This report focuses on ship detection based on satellite Synthetic Aperture Radar (SAR) images. Imagery from the European ENVISAT satellite is analysed and Automatic Identification System (AIS) data from the Norwegian Coastal Administration (Kystverket) is applied for verification. Aegir is a tool for analysis of SAR images developed at FFI. Aegir contains a module for ship detection. Two different thresholding algorithms for automatic ship detection, N-sigma and K-distribution, are implemented in Aegir. The algorithms are tested on ENVISAT ASAR Alternating Polarisation (AP) images. The results from these experiments are presented in this report. The results show that the total number of alarms generated for co-polarised channels (VV and HH) is larger than the total number of alarms generated for cross-polarised channels (VH and HV). This result is independent of the choice of algorithm: N-sigma or K-distribution. For the N-sigma algorithm, we found a higher false alarm rate for VV and HH than for VH and HV. A slightly higher number of verified ships were also detected in cross-polarisation. Comparing the N-sigma algorithm with the K-distribution algorithm, we found that the K-distribution algorithm produces a lower false alarm rate both for cross-pol and co-pol data, compared to the N-sigma algorithm. However, slightly fewer verified ships were detected with the K-distribution algorithm compared to the N-sigma algorithm, both for cross-pol and co-pol. In this context, it is important to remember that “false” alarms can be vessels without their AIS transponder turned on. In the case of a moving vessel, there could also be a time difference between the AIS reports and the SAR acquisition, making a match between the AIS position and the position in the SAR images highly uncertain. In addition, a ship which is heading in a non-azimuthal SAR direction will cause an azimuth displacement in the image, due to Doppler shift effects. This should also be taken into account when matching the SAR ship position with the positions found in the AIS reports [6] [12]. Further work should include a more detailed study of the causes of the false alarms. The algorithms should also be tested on a larger data material where scenes with varying sea state, incidence angle and polarisation are analysed. The ship detection system described in this report is not yet mature enough for an operational context, and the results presented here must be viewed as preliminary. In the future, the K-distribution algorithm will be tested on coherence images (or Internal Hermittian Product (IHP) images) from Norut AS. The coherence images are generated based on single look complex data. The idea of such an approach is to utilize the higher phase coherence of a vessel than the surrounding sea to enhance the vessel signature and to suppress the sea surface signature [8]. The applicability of the K-distribution to model the sea clutter in these images still needs to be evaluated.