Add OpenEarthMap-SAR - DFC2025 baseline U-Net weights

[isaaccorley]

Adds the DFC2025 Track 1 Baseline U-Net weights trained by @cliffbb in the baseline repo. The model is a land cover land-use semantic segmentation model trained on the OpenEarthMap-SAR dataset which contains high-resolution 1-channel uint8 SAR images from Umbra Space
particularly the Geo-Ellipsoid Corrected (GEC) geotiffs.

The model has an IoU of 0.41 on the private test set leaderboard

Before @calebrob6 asks, I've checked that it runs and it's decent but seems to get confused by the noisiness in the SAR imagery. Below is a minimum working example:

import random
import glob

import torch
import torchvision.transforms.v2.functional as F
import rasterio
import matplotlib.pyplot as plt
from matplotlib.colors import to_rgb, ListedColormap
from torchgeo.models import Unet_Weights, unet

h = w = 1024
images = sorted(glob.glob("/data/train/sar_images/*.tif"))
labels = sorted(glob.glob("/data/train/labels/*.tif"))

device = torch.device("cuda")
dtype = torch.float32
weights = Unet_Weights.UMBRA_GEC_OPENEARTHMAP_SAR
classes = weights.meta["classes"]
colormap = ListedColormap([to_rgb(hex) for hex in weights.meta["colormap"]])

transforms = weights.transforms
model = unet(weights)
model.eval()
transforms.eval()
model = model.to(device).to(dtype)
transforms = transforms.to(device).to(dtype)

idx = random.randint(0, len(images) - 1)
path = images[idx]
label = labels[idx]
with rasterio.open(path) as src:
    image = torch.from_numpy(src.read())
    image = F.center_crop(image, (h, w))
    x = image.clone().unsqueeze(dim=0).to(device).to(dtype)
    image = image.repeat(3, 1, 1)

with rasterio.open(label) as src:
    mask = torch.from_numpy(src.read())
    mask = F.center_crop(mask, (h, w))
    mask = mask.squeeze(dim=0).numpy()

x = transforms(dict(image=x))["image"]
preds = model(x).cpu().squeeze().argmax(dim=0)

fig, axs = plt.subplots(1, 3, figsize=(12, 5))
axs[0].imshow(image.permute(1, 2, 0).numpy())
axs[0].axis('off')
axs[1].imshow(mask, vmin=0, vmax=len(classes), cmap=colormap, interpolation='none')
axs[1].axis('off')
axs[2].imshow(preds, vmin=0, vmax=len(classes), cmap=colormap, interpolation='none')
axs[0].set_title('Image')
axs[1].set_title('Mask')
axs[2].set_title('Predictions')

[Copilot]

Pull Request Overview

This PR adds DFC2025 Track 1 Baseline U-Net weights for land cover segmentation using SAR imagery from the OpenEarthMap-SAR dataset. The model achieves 0.41 IoU on the private test set and is trained on high-resolution Umbra Space SAR data.

• Adds new UMBRA_GEC_OPENEARTHMAP_SAR weight configuration to the U-Net model
• Updates documentation to rename "Sentinel-1" section to "Synthetic Aperture Radar (SAR)" for broader SAR model coverage
• Includes model metadata with 9 land cover classes, colormap definitions, and encoder specifications

Reviewed Changes

Copilot reviewed 3 out of 3 changed files in this pull request and generated 1 comment.

File	Description
torchgeo/models/unet.py	Adds new SAR weight configuration with metadata and model kwargs support
docs/api/weights/sar.csv	Adds entry for the new U-Net SAR weights in documentation table
docs/api/models.rst	Updates section title from "Sentinel-1" to "Synthetic Aperture Radar (SAR)"

[calebrob6]

I see my job here is not needed

[adamjstewart]

I disagree with this change. There are many different kinds of SAR (C-band vs. L-band) just like there are many different kinds of optical imagery (Sentinel-2, Landsat, etc.). Just because two models have the same number of bands doesn't mean they will be compatible with any SAR imagery.

[isaaccorley]

That's okay with me but you're not proposing a solution here. Do you prefer that I just make a separate group for "Umbra Space" in this case? Because this isn't Sentinel-1.

[adamjstewart]

Yes, separate group/table for Umbra

[adamjstewart]

Let's move sorting to a separate PR to make this easier to review.

[adamjstewart]

This is a lot more metadata than other models. Is this coming from STAC MLM?

[isaaccorley]

Most of our weights are foundation models with no class specific outputs. I think it makes sense to add this here for the user to have information about the output order and categories.

[adamjstewart]

B1 doesn't sound right, is it VV or VH? These are very different polarizations, an incompatible choice could result in the poor performance you are seeing.

[isaaccorley]

It's a different level of processing that is used from Umbra where it converted to a grayscale uint8 image. This doesn't use the raw band data unfortunately.

[adamjstewart]

Where can I find this license documented?

[isaaccorley]

The dataset is CC-BY-4.0 licensed so the model is also the same license found here https://registry.opendata.aws/umbra-open-data/

[isaaccorley]

@cliffbb can you clarify here what license you are releasing the model weights under?