Does the 4th band (alpha band) matter when using RGB layer as input for pre-trained 3-band Deep Learning model?

Question

I would like to use one of ESRIs pre-trained deep learning models. It specifies that a 3-band RGB layer should be used as input. However, my RGB layer that I plan to use as input contains a fourth layer, an 'alpha' layer. Do I need to remove this layer in any way before using it as input for the deep learning model? Or can I just use the layer as it is, and trust that the model will work fine with it?

bbooth · Accepted Answer

The imagery you use should match the imagery the model was trained on.

You can use a raster function to make a version of the imagery without the fourth band.

This tutorial shows the process:

https://learn.arcgis.com/en/projects/improve-a-deep-learning-model-with-transfer-learning/#:~:text=using%20a%20raster%20function