Model predicting null class at edges of chips

[NoaMills-USDA]

I've been tinkering with the quickstart model, and I get null pixel predictions in areas where data is indeed present. When I run the quickstart as is, I get null predictions on the borders of the validation raster where the chips don't reach. I made one change to the quickstart: I changed the pixel size from 300 to 325 so the prediction covers the entire area of the validation raster. Now, I get null pixels predicted in the middle of the raster, especially at the corners of the chips. Why is this? I'm only training for one epoch, but I figured there shouldn't be any guesswork in which pixels have no data.

Quickstart validation prediction:

Quickstart validation prediction after changing chip size to 325:

[AdeelH]

To the model, the null class is just like any other class. If you're only training for 1 epoch, it is likely that the model has just not learned to recognize no-data pixels yet.

The null class pixels in the prediction are from two sources:

• the model's predictions
• RV automatically setting no-data pixels to the null class