Composite Raster in Deep Learning

[Ed_]

Hello,

Let's say there are 4 different types of rasters (three are DEM derived and one is NDVI), A, B, C and D, and use the `Composite Raster` tool to create a composite raster from these four rasters, that will later used to train a Deep Learning model and again used as input raster in `Classify Pixels using Deep Learning`.

Does the order of raster variables in the composite raster make any difference when using Deep Learning?

Pro version 3.4.2 and model in question in PSPNet and UNet.

[DanPatterson]

Did you see something in the PSPNet and UNet documentation to suggest that there was?

How U-net works? | ArcGIS API for Python  for example

Not wanting to go down the documentation rabbit hole. Or is this just a "just wondering" question,

[Ed_]

Hi Dan, hope all is well, yes it's a wondering question. 

[Ed_]

@PavanYadav @ShivaniPathak any clue on this please?

[PavanYadav]

Hi @Ed_ Please maintain the same order as the training data. For example, if the band combination in the training data was raster1, raster2, then ensure the input raster for inference follows the same order: raster1, raster2.

[Ed_]

Hi @PavanYadav thank you for the response, so the order of band combination doesn't matter much in the whole process? For example if deep learning process is done first with band combination for training and inference was raster 1,2,3 and 4 and then another iteration of training and inference was done band combination of raster 3,2,1 and 4. The results would still be same/similar?

[PavanYadav · Accepted solution]

hi @Ed_ if you trained a model with 1,4,2,3; for inferencing you will need have the same combination. For any reasons if you need to retrain this model; you will need to create training data with the same combination e.g. 1,4,2,3.. 

But, correct me if I am mistaken -- I guess you're asking if there would be any difference if you have 1,2,3,4 or 3,4,1,2? I cannot say for sure because it's something I have not investigated much. Theoretically, it may impact the learning.. deep learning models, particularly CNNs, handle each band as an individual 'channel.' The model learn to identify patterns and features depending on the specific sequence of these channels. Consequently, the way the model interprets the data might be directly affected by the order in which the rasters are arranged. You might want to experiment with different combinations to see which one yields better results