geohai/sled-s1-s2-ls

This model has been pushed to the Hub using the PytorchModelHubMixin integration:

• Code: https://github.com/geohai/sled
• Paper: [More Information Needed]
• Docs: https://github.com/geohai/sled

This version of SLED embeds in 768 dimensions and uses the RFF position encoder popularized by GeoCLIP. It is trained on

• 40k UAR Sentinel-2 L2A images from the S2-100K dataset
• 40k random Landsat 8/9 OLI_SR images from the SSL4EO dataset
• 20k UAR Sentinel-1 VH/VV polarizations, downloaded directly from Microsoft's Planetary STAC API

It trains against the following teacher encoders

• Terra-FM for Sentinel-2 imagery
• TorchGeo's Landsat_OLI_SR MOCO model for Landsat 8/9 imagery
• FG-MAE for Sentinel-1 imagery

To load these weights, run the following:

from sled_geo import get_encoder
my_sled_model = get_encoder.get_rff_encoder(embed_dim=768)
my_sled_model.from_pretrained("geohai/sled-s1-s2-ls")
new_york_location = torch.tensor([[-73.935242, 40.730610]]) #lon,lat
my_sled_model(new_york_location) #returns a tensor of [1, 768] shape