Experimental model for segmenting vegetation on satellite images.
And that is it. It just labels pixels as "vegetation" OR "other".
Created by finetuning the facebook/maskformer-swin-base-ade model, and training with a small number (~25) of manually labeled satellite images of urban-ish areas.
BIAS WARNING:
This model was created for a personal art and urbanism project and while the training set included images from geographically diverse cities of personal importance to me, it is in no way exhaustive. There are no cities in Asia, Africa, Central America or Oceania.
The urban areas included were of, or around, these cities:
- São Paulo, BR
- Rio de Janeiro, BR
- New York, US
- Pittsburgh, US
- Oakland, US
- Berlin, DE
- Milan, IT
- Riyadh, SA
EVALUATION WARNING:
Anecdotally speaking, it seems more precise than the original facebook/maskformer-swin-base-ade model model when used to get masks for vegetation.
It works with images of other cities, but success criteria is qualitative.