Unclassified (Class 0) Pixels in Deep Learning Output Raster

[GBacon]

I produced some very good results when using the "High Resolution Land Cover - USA" deep learning package with NAIP imagery. There is a relatively small number of unclassified (0) pixels in the raster output that I'm wondering if others have seen. In most cases, this seems to happen in areas where tree canopy is over a road or other impervious surface as shown below. I can easily fix this with the Image Analyst pixel editor, but wondered if this could be prevented somehow since it happened in two different years of imagery, one natural color and one IR. Has anyone else seen this? I used ArcGIS Pro 3.1 and associated deep learning libraries.

Impressive results though!

[JeffreySwain · Accepted solution]

Hey @GBacon ,

Did you use the default settings with the deep learning package?  If you are missing some areas, then you might want to try to change to the less detailed classes to see if you get the complete coverage. That particular package doesn't have any threshold to adjust, but just remember those model parameters are there for you to adjust to fit your data. So try altering them to see if they improve the results of your classification. 

[GBacon]

I just tested a small area that had this problem, and the results look much better with the simplified classes. Thank you!

[JeffreySwain]

Great ! All of the deep learning models have parameters to adjust to meet your classification needs.