I suggest you use https://github.com/opengeos/segment-geospatial instead of this repo. It was the first successful attempt to join SAM and EO data, but now there are much better documented and much better maintained options.

Earth observation tools for Meta AI Segment Anything

Facebook Research Segment Anything — Apache-2.0 license

Gumblex tms2geotiff — BSD-2-Clause license

aeronetlib — MIT licence

Other code — MIT license

Segment Anything and tms2geotiff were copied to this repo 9 Apr 2022, you can update them to more recent versions if needed

• download TMS data (including OpenAerialMap and Mapbox Maxar) as GeoTIFF files
• process GeoTIFF files with Meta AI Segment Anything models
• save predicted segments as GeoTIFF raster data and GPKG vector data

and a little bit more

• jupyter notebook in the repo https://github.com/aliaksandr960/segment-anything-eo/blob/main/basic_usage.ipynb

• Using a sliding window algorithm to process large images
• In order to separate instances, every instance gets surrounded by 1px width spare space, so it is not the same as how original Segment Anything works

Segment Anything was released less than a week ago, and these are the first experiments with it. I don't know how paramters affect perfomance — feel free to change everything.