I'm trying to train a model to predict landslide risk using DL. I'm not going to use the build-in tools in ArcGIS pro to train my model as I want to tweak the network myself. The problem is I'm not sure if I should just merge all my classification layers into one new raster and use it to train the model as it is, or do something similar to 'encoding' for my classified layers. I'm rather new to CNN models.
For example, let's say I have a vector layer with land use classification and there are 3 land uses in the vector dataset. Do I create 1 raster that has values "1, 2, 3" (or something similar) in its pixels representing every 3 land use, or do I create 3 rasters with pixel values "0, 1" where 1 means presence of a class and 0 means absence of it? If encoding is the solution, is there any tool in DL libraries to deal with this problem in 2 dimensions?