Interesting, any idea why SAM tends to under-segment in some areas?

I have no idea but I noticed those areas tend to have a sandy color with less contrast between plots.

Also, looks like the "context" of the whole block influences the segmentation in some way, for example, if there's a big pond and and a big forest patch the surrounding plots are not segmented. These are only my first impressions after a quick review of the results.

The size of the blocks/tiles has a huge impact on sam segmentation, that's the main conclusion after spending some time seeing different results with Sentinel 2

Interesting, probably a good feature to adapt to different image compositions of photographs but it does not sound ideal to get a consistent segmentation over the full extent of satellite imagery. Thanks for sharing!

@andrestumpf
Yep, not ideal. We're thinking on adding some kind of "tile stitching" mechanism but it isn't trivial.