How do I deal with tiles in the Orfeo Toolbox segmentation?

Question

I'm using the Orfeo Toolbox for segementation of large Very High Resolution Satellite images. It performs very well and i'm mostly very happy with it. But I still get some weird/bad results that I assume are related to the partition of the process in tiles.

I'm doing this:

#Segmentation variables spatialr = '7' maxiter = '6000' ranger = '2,6' minsize = '8000' tilesize = '1048' simplify = '0.4' # SEGMENTATION
print "Performing image segmentation" app = otbApplication.Registry.CreateApplication("Segmentation") app.SetParameterString("in", imagem_base_com_ndvi) app.SetParameterString("filter.meanshift.spatialr", spatialr) app.SetParameterString("filter.meanshift.maxiter", maxiter) app.SetParameterString("filter.meanshift.ranger", ranger) app.SetParameterString("filter.meanshift.minsize", minsize) app.SetParameterString("mode.vector.inmask", image_mask) app.SetParameterString("mode.vector.stitch", 'True') app.SetParameterString("mode.vector.out", vector_out) app.SetParameterString("mode.vector.simplify", simplify) app.SetParameterString("mode.vector.outmode", 'ovw') app.SetParameterString("mode.vector.tilesize", tilesize) app.SetParameterString("mode.vector.neighbor", 'True') app.ExecuteAndWriteOutput()

Here are some examples of the resulting segmentation:

Anyone has ideas to help me avoid these errors?

Thanks.

Answer 1

To better understand, I have done a similar segmentation in GRASS GIS 7 using the new i.segment. In my opinion these lines appear where the image data were mosaiked due to a non-perfect histogram matching (or whatever) being used.

In short: Orfeo or any other segmentation software may deliver better results when the initial mosaiking is improved to avoid the borders being visible. This may require a seamline mosaiking approach.

Proof: Using your data as provided in your comment, I performed the following steps (reported here for reproducibility):

Creating a GRASS GIS location from the dataset:

gdalinfo test.tif
# -> it is UTM36S
grass70 -c test.tif utm36s

Then data registration (rather than true import) of the GeoTIFF:

r.external test.tif out=image_to_be_segmented

Setting the computational region to the map:

g.region rast=image_to_be_segmented.1 -p

Then image segmentation (note, that I used simply the default settings!) and vectorization:

i.segment image_to_be_segmented out=segmented_image threshold=0.5
r.to.vect segmented_image out=segmented_image type=area

Eventually visualize the results:

d.mon wx0
d.rgb b=image_to_be_segmented.1 g=image_to_be_segmented.2 r=image_to_be_segmented.3
d.vect segmented_image type=boundary -c

Full sample map (as you see I did not clean up small area with v.clean:

Zoomed result (here the mosaiking problem is well seen which results in straight, undesired lines, you need to look at the full extent uploaded image):

For a final result, small areas should be removed etc but in the first place proper input data are needed.

Answer 2

note that next OTB release (3.20) will provide a new segmentation workflow based on the mean-shift algorithm which allows to perform tile-wise segmentation of very large image with theoretical guarantees of getting identical results to those without tiling.

You can find a description of the workflow (different from the one cited above) here.enter link description here

HTH