Wetland Identification Model - Train Random Trees output with all bands' importance as zero

[MalloryGill]

Hey all - I'm running the ArcHydro tools in Pro (2.9.3) and using the wetland identification toolset. I'm getting stuck at the Train Random Trees tool using DTW, Curvature, and TWI as predictor inputs - the predictor input rasters have the expected value ranges and look how I'd expect them to. The tool runs "successfully" but I get a warning that says "Extents between the training raster and predictor variable raster do not match. The predictor variable(s) will be clipped to the training raster extents temporarily. Only cells overlapping the cells in the training raster can be used to train the model." But the extents of the predictor variables all appear to be the same (and are all built using the other WIM tools that derive these predictor inputs from the same origin DEM file.) 

In the variable importance txt file, all the band importances are 0.0, and the bands listed  for the raster in pro are labeled as are 1, 2, and 1 (instead of the expected 1, 2, and 3). I've tried creating a fresh APRX with no improvement and I'm not sure what else I can try to get it working. Any ideas what might be causing this?

Thanks!

[GinaO_Neil]

Hi Mallory,

Thanks for your question. Happy to hear you are using WIM! The warning you are seeing in Train Random Trees is expected. I may just downgrade that to a message rather than a warning. It is just notifying you that the training raster does not overlap perfectly with the predictor variables. Typically this happens because predictor variables are calculated for a catchment, but training data only encompasses a subset of that catchment. So the tool does some slipping internally and the model only trains using the pixels where training raster overlaps with the predictor variables. This does not indicate the extents mismatch between the individual predictor variable rasters. 

The band importance measures of 0 are strange. I can imagine that happening if there are no training cells that align with predictor variable cells. Can you share a screenshot of the training raster and predictor variables?

Thanks,

Gina

[MalloryGill · Accepted solution]

Hey Gina - thanks for the reply! It turned out to just be user error with the projection on one of the input rasters. All the raster tifs aligned almost by definition since I'm using the same small grid square to clip all of them down to the same cell size/extent. And I'd checked all of the projections to make sure they were they same but managed to overlook that one of the rasters was a very similar projection but not actually the same. I only caught it when I was going through all the "checks" again and was comparing the WKID numbers and noticed it was 1 off from the rest. Once I got that sorted out it runs as expected (still with the extents warning message) and successfully assigns band variable importances and generates outputs I'd expect! 

[GinaO_Neil]

Glad to hear it! Thanks for following up.