How can I increase boundaries accuracy in satellite images?

[erabdi78 (Erfan Abdi)]

I use the following codes for this work, but they have low accuracy. I would be grateful if you could help me.

The codes are as follows

Pruning@Thinning@Closing[#, 5] &@DeleteSmallComponents[#, 500] &@
     LocalAdaptiveBinarize[#, 2] &@GaussianFilter[#, 10] &@img

ridges = ImageAdjust[RidgeFilter[imge, 7]]

bin = MorphologicalBinarize[ridges, {.1, .2}]

dist = DistanceTransform[ColorNegate@bin];
maxMarkers = MaxDetect[dist, 10];
HighlightImage[bin, maxMarkers]

watersheds = WatershedComponents[ridges, maxMarkers];
Colorize[watersheds]

[davitb (Davit Baghdasaryan)]

Hello,

One way to improve the accuracy, in my opinion, is to train, say, Unet neural network (see how to do that here ) on some data, for example, public dataset LPIS 2016 for semantic segmentation and then use WatershedComponents similar what you did to go to instance segmentation.

[markusa (Markus van)]

Hello Erfan,

EdgeDetect[img, r] (with r = 2, 3 or 4) does a good job finding boundaries. It’s sensitive and it has a good location accuracies.

However, if you are seeking a good and consistent segmentation of all the fields and roads try ClusteringComponents[img].

Training a neural net may eventually return better results, but it takes quite some effort.
We are working on next generation foundation networks to provide you better solutions in upcoming versions.

[erabdi78 (Erfan Abdi)]

Hello Mr. Markus, Thank you for your guidance.