Segmentation Accuracy

Hi all,

I am experimenting with the segmentation options in SNAP (e.g. Batz&Schape - or Lambda like the FLS in ERDAS) and all seem to be working fine - but what I would like is a way of quantifying results under different settings, using methods you see in literature such as oversegmentation, undersegmentation, area fit - etc.

Firstly - which which would you prioritise?
Secondly - what is the best way of doing this in SNAP?

image

Thanks,

Conor

I don’t have a solution (or an opinion) but may I ask what you do with the segments in SNAP - as these are just raster patches representing the number of each segment.

I saw that mentioned that in an older post when i searched earlier today - but hoping it had been clarified in later versions as to be honest I am not sure what happens next.

The user needs to then classify those rasterized segments but I cant find an attribute table to see what it is at currently. I was looking at this in Scikit a while back and found a good tutorial:

First you need to assign them a unique identifier attribute

Then you classify those using SVM or RandomForest or similar - so somehow this output needs to feed into the RF classifier in SNAP, ‘train by raster’ and ‘train by vector’ are an option but thats the training data i presume, not classification datatype input.

1 Like

Thank you. It would indeed by great to have the segments as vectors whose attributes could at least be exported. My question was rather adressing if you already found a way to use these segments within SNAP.
Still, thank you for the reference.
In the past I had my own workaround used segments generated in QGIS and calculated zonal statistics, then I exported the attribute table to analyze it in Orange (https://orange.biolab.si) and imported the classified table back into QGIS to join it there with the geometries. So basically what the python script does.

Sounds complicated!

I tried modelling the whole process in QGIS with the OTB add-on in the model builder - but one of the later steps kept crashing on me, not sure what the issue was and got sidetracked.

No wonder eCOGNITION and Imagine Objective are so popular!

1 Like

not so complicated. I like the visualization tools in Orange and the possibility to train and apply several classifiers in on e run.

if you still seach an open-source solution for the full workflow you might try this one: https://github.com/ANAGEO/Opensource_OBIA_processing_chain

1 Like

Must give it a look - thanks for the info!