I have a small (processed SAR image, RGB) area of 40 * 40 km approx. where I want to extract water bodies (it's an estuary during flooding basically), i.e. apply pixel classification (I assume object class. isn't for this application).
1. Export training data - I select 20 samples for water and 20 for wetlands, randomly over area, no issues with export (mask used to reduce the area), selected classified tiles
2. Train data for deep learning - using the aforementioned exported data and U Net model (tried several cases up to 60 epochs with different batch sizes, my GPU has 8 GB and that shouldn't be an issue).. I try to train the model but results are showing 0 good predictions and 0 pixels classified in my selected classes.
3. Classified pixels using deep learning - since previous steps failed, makes no sense to start this yet...
I am relatively new in deep learning and wondering does it even make sense to apply it while studying only 1 single satellite image ?
I can do this task very easily with the same samples with traditional image classification, but still don't get why deep learning isn't working.
Attached:
-Image with samples
-Train results
Cheers
Try reducing the epochs to 20 or less and see if that helps. I have a 4 GB RTX GPU (little fella has to do a lot of hard work lol) and I never went beyond 8 batch size.
thanks but think I can say I've already tested all scenarios regarding these parameters (epochs, batch size, ..), it must be something with e.g. sparse sampling or ArcGIS ...
Yeah maybe that sample size does seem quite small but then again you know your study area better to decide on the sample size. For example in my last case of trees my training sample size included hundreds of trees. However, if increasing the sampling does not help either then maybe try reinstalling the ML libraries.
this was enough to perform both unsupervised and supervised classifications successfully though ... Generally, looking at amount of questions and reported errors on deep learning arcgis package it seems that developers still have work to do on it...
Well ArcGIS Pro 3.1 Deep Learning works perfectly fine in my case but yes your case might be different, good luck.