SAR ship detection in dual polarization combination channels
About the publication
Report number
18/00886
ISBN
978-82-464-3065-2
Format
PDF-document
Size
2.4 MB
Language
English
Satellite Synthetic Aperture Radar (SAR) imagery is widely used for automatic ship detection.
SAR data comes in different modes and polarization combinations. Dual polarization ScanSAR
products offer good spatial coverage, and are thus often the mode of choice. Dual polarization
SAR products consist of a co-pol and a cross-pol image. Utilizing the full potential of the data is
an important aspect in any surveillance task. In this study, two dual polarization channel
combinations are presented and evaluated for use in automatic ship detection. The first one,
named dual, is a multiplication of the individual polarization amplitudes, scaled with an
estimated average sea background. The second combination channel, named dual-int, is a
multiplication of the individual polarization channel intensities. Both channels aim to increase the
ship-to-sea contrast in the SAR data, as one of the main objectives is to detect bright pixels in
the images. Data from two platforms, RADARSAT-2 and Sentinel-1, have been investigated in
this study.
A short state-of-the-art literature review on dual polarization channel combinations in automatic
ship detection has been done to put the presented methods in context.
The statistical properties of the different channels used in the automatic ship detection have
been investigated. The RADARSAT-2 data fits well with the proposed probability density
functions (PDFs), while the Sentinel-1 data fits the PDFs to a lesser degree. Recommendation
on a more in-depth analysis of the statistics of Sentinel-1 SAR data is made.
Results from the automatic ship detection show that both dual polarization combination
channels generally outperform the co-pol channel in terms of identified ship detection counts.
Comparison with the cross-pol channel shows similar performance for both presented dual
polarization combination channels. When looking at the added value provided by the proposed
channels, results differ between the two platforms. For RADARSAT-2, both the dual and the
dual-int channel offer added value in terms of an increase in the number of identified detections.
This is not the case for Sentinel-1. However, the detections can still be used in an updated
confidence estimate procedure for both platforms.
Analysis of the false alarm rate results shows that the combination channels both perform worse
than the cross-pol channel. Two special cases in which the co-pol channel is known to produce
many false alarms have been investigated. At low wind conditions, patterns in the SAR co-pol
image due to algae growth can produce high false alarm rates. It is shown that this problem
does not seem to persist in the combination channels. Secondly, false alarms at low incidence
angles are common in the co-pol channel. Both combination channels have fewer false alarms
than the co-pol channel, but the counts are still very high.
Comparing the two presented combination channels, the dual channel comes out best and
recommendations are made to use this channel. Also, if timing is crucial, it is recommended
only to process the cross-pol and the dual channel for RADARSAT-2 imagery, and only the copol
and cross-pol channels for Sentinel-1 imagery.