Are there any 4-band backbone models for the Train Deep Learning Model tool or is there a way to use my own backbone?

[Ruari_Langland]

Just wondering if anyone has any knowledge of doing deep learning on 4-band imagery using the Arc Pro built in tools?

Additionally wondered if anyone had any luck applying a "bring your own model" approach to setting the Backbone Model in the tool. There are some great models on huggingface I would like to try, and they may need modification to accept a 4th channel:

https://huggingface.co/timm

Thanks in advance!

Ruari

[Ed_]

@PavanYadav i have the same question, can you please share some resources on this? thank you!

[PavanYadav · Accepted solution]

Hi @Ruari_Langland  and @Ed_ 
You can find some examples at https://developers.arcgis.com/python/latest/samples. This is one of them: https://developers.arcgis.com/python/latest/samples/landcover-classification-using-hyperspectral-ima... 

The following is from a project I did sometime ago.. this one is 4-bands: i used RESNET34 .. the results were good 

Here are my training parameters 

max_epochs = 25
model_type = "UNET"
batch_size=8
arcpy.ia.TrainDeepLearningModel(in_folder=[TrainingSampleForLandcoverClassification],out_folder=Model_LandcoverClassification, max_epochs=max_epochs, model_type=model_type, batch_size=batch_size,arguments=[["chip_size", "224"], ["resize_to", ""], ["monitor", "valid_loss"]], backbone_model="RESNET34",validation_percentage=10, stop_training="STOP_TRAINING", freeze="FREEZE_MODEL")[0]