putvision / landcoverseg Goto Github PK

Hi, thank you for uploading this. I have successfully trained and exported the included model on our university cluster and have tested that it works in my Deepness plugin.

I am starting to create my own custom dataset and am using the following as my image and mask (actual resolution 4284x4218):

The colors in the mask are arbitrary single byte values I chose when generating the mask (1, 150, 200) as I have 3 classes (I think

assumes 5 classes so my class configuration in the yaml has two extra classes, ['background', 'shrub', 'forest', 'null1', 'null2'])

I trained and it ran for 148 epochs before finishing. I exported and tested out the model on the training area and got the following (overlaid on my training mask for display here):

It looks like it is assigning channel_0 to most of the image, and there is some assignment of channel_1, presumably my shrub layer. None of the forest layer (value 200) were assigned. I tried a larger extent as well, and only got assignments from the first 2 layers.

I can produce much more training data, but wanted to make sure I am using the software correctly first. When defining classes: ['background', 'building', 'woodland', 'water', 'road'], how do you know which pixel values in the raster these classes will be mapped to, is it based on byte order?

And do you think my lack of any forest assignments is based on my low amount or low quality of training data (e.g. my masks for the forest layer are too messy), or maybe it is something I am configuring incorrectly?