Combining Multi-Variate Statistics and Speckle Reduction for Line Detection in Multi-Channel SAR Images

The current paper focuses on the detection of linear objects (e.g. roads, rivers, tree lines, etc.) in multi-frequency polarimetric SAR images. We obtained sets of full-polarimetric P-baud and L-band and VV-polarised C- and X-band images. The images cover the same region but have a different spatial resolution. We also obtained transformation matrices that relate the slant-range coordinates to geocoded coordinates for each frequency band. The detection of linear features is performed on each of the slant-range images and the results are then geocoded and fused. In SAR images, for deciding whether a line passes through a given point, a relatively large neighbourhood has to be considered because of the speckle. Normally a set of rectangles is scanned over the image and at each point the statistics of the pixels inside the different rectangles are compared to decide whether a line is present. For single-channel data, a line detector is constructed from the Touzi edge detector. For polarimetric data, we use a multi-variate hypothesis test. Because of the difference in spatial resolution and information content of the 4 frequency bands, results are improved by fusing the individual results from the different bands. On the other hand, the synergy with speckle reduction is also examined. Without speckle reduction, large scanning rectangles need to be used for the line detection because of the presence of the speckle. If speckle reduction is applied prior to line detection, smaller rectangles can be used. The former approach allows to detect lines that show a lower contrast while the latter allows to find smaller details and achieves a higher spatial accuracy. The proposed method was applied to one of the sets of images and results are shown and evaluated.