Using feature class as a 'segmented' reference to collect training samples for different image

[Anonymous User]

I have collected training samples for deep learning in ArcGIS Pro using the Training Samples Manager for True Color combination. The process was done with the aid of a segmented layer developed for the True color image. But I would like to be able to use the vector/feature class generated from this True color training samples collection as a reference to collect training samples for a similar image with different band combinations; False Urban. 

I understand that the training dataset feature class generated by Training Samples Manager has established its own RGB values, but can't the vector itself be utilized as a substitute for a segmented layer that can help with the training dataset sampling; just with a different band combo?  

The endgame is to export different training datasets for different band combinations to be separately used in training the deep learning model. 

Any pointers is very much appreciated! 

Lea

[AmlanR · Accepted solution]

Hi @Anonymous User . Assuming you already have a feature class with a column for the class value, you can definitely use the same feature class on multiple imagery to collect separate set of samples. 

You can do so by using the 'Export Training Data for Deep Learning' GP tool (Image Analyst)

Export Training Data For Deep Learning (Image Analyst)—ArcGIS Pro | Documentation

If the features were collected from training samples manager (for machine learning) or using Label objects for deep learning (for deep learning), you should be able to save a copy of the feature class and use it again.

If you have a feature class with the RGB values in it, creating a copy of the feature class and deleting the RGB and blob fields will work just fine.

[Anonymous User]

Hey there Amlan!

Sorry for the delay. I've tried your suggestion and it worked! Although it is not necessary to remove the Red, Blue and Green columns. I guess, for the sake of exporting the data for deep learning model training is really just referring to chunking them out into image chips. 

I've tried out all the way to training the model and classifying the pixels and it works without any issues; skipping over horrendous misclassifications that I believe is my error. 

Thanks again!!!