segments.htm

ruta

A georeferenced Landsat TM image over Kalmar in the south of Sweden, using bands 7,5,4.

The processing steps leading up to the image on the left were:
-A grayscale image was derived by averaging the 3 bands. This can be done in more refined ways but usually won't affect the result.
-A Gauss filter was applied to yield a gradient image.
-Standard Vincent-Soille watershedding segmentation, removal of shed-barriers by assigmnent to closest pixelvalue. As can be seen this results in massive over-segmentation. The animation shows the steps in the region merging
The following windows bat-file did the job. The files "parvec0,parvec1,parvec2" contain the actual parameters, and examples can be found in the Application directory in the Imlib software distribution

@ECHO OFF
REM A script for segmentation and region growing. Assumes the existence
REM of 3 parameter files in the same directory
REM
REM
IF "%1"=="" GOTO usage
:process
segmerge %1.ppm nofile nofile %1-seg.bil %1-edg.ppm parvec0
segmerge %1.ppm nofile %1-seg.bil %1-seg2.bil %1-edg2.ppm parvec1
segmerge %1.ppm nofile %1-seg2.bil %1-seg3.bil %1-edg3.ppm parvec2

goto end

:usage
Echo goto end

:end
::

A color composite of ERS-JERS-ERS images, extracted from a test area outside Uppsala (NOPEX).
Red colour is June image, green May, and blue April. No speckle filter has been applied.

rada-edg Result after a first and a second region merging pass using the same T-test and roundess criterion as for the landsat image above. A Gauss filter with large standard deviation was used to compensate for the speckle.