How does one create land cover/use classification maps containing more than three classes without any gaps

[paulvpop]

Hello everyone, I have developed a free and open source end-to-end land cover classification workflow, which I have documented in both GitHub and Open Science Framework. I use the FORCE modules for a significant chunk of the workflow with some customisation.

While the workflow itself is largely complete (with the documentation for some sections yet to be updated), it doesn't produce gap-free maps in land cover maps. I feel like I am doing something wrong because I am getting land cover maps with more than 70% gaps/unpredicted pixels after individually producing maps for each class and then integrating them. Is this a flaw in the algorithm that FORCE uses or is there some issue in my workflow? Could someone check and let me know?

The issue likely lies within this section: https://github.com/paulvpop/gis-land-cover-mapping/blob/main/07.%20Sampling.md#step-a1-create-sampling-parameter-file

With kind regards,
Paul Pop

[davidfrantz]

Dear @paulvpop.

I recall having answered this issue before. Nevertheless, the problem is that you are splitting the samples into individual files, followed by training individual models for each class.

The solution is as simple as skipping Step B and using the output of Step A directly in 08. This only trains one model and gives you a multi-class classification.

Regards,
David

[paulvpop]

Thank you David. I had issues with the straight forward method when I last tested it (which prompted me to create Step B), but now it's working properly! Thank you. I will update the workflow