umbra_gec_unet_effb4_openearthmap_sar

This repo contains the baseline model for the 2025 IEEE GRSS Data Fusion Contest Track 1 which is trained on the OpenEarthMap-SAR dataset for land cover land use semantic segmentation.

The model is trained on SAR imagery from the Umbra Open Data Program which is CC-BY-4.0 licensed.

See for details: github.com/cliffbb/DFC2025-OEM-SAR-Baseline

Model weights extracted below:

import os
import hashlib

import torch
import gdown
import segmentation_models_pytorch as smp


id = "1Myd8b2KVFRuYVPyjB6EAv70OsNmjtgB9"
gdown.download(id=id, output="model.pt")
state_dict = torch.load("model.pt", weights_only=False, map_location="cpu")
in_channels = state_dict["encoder._conv_stem.weight"].shape[1]
classes = state_dict["segmentation_head.0.weight"].shape[0]
print(in_channels, classes)

model = smp.Unet(
    classes=classes,
    in_channels=in_channels,
    encoder_weights=None,
    encoder_name="efficientnet-b4",
    decoder_attention_type="scse",
)
model.load_state_dict(state_dict, strict=True)
filename = "umbra_gec_unet_effb4_openearthmap_sar.pth"
torch.save(model.state_dict(), filename)
md5 = hashlib.md5(open(filename, "rb").read()).hexdigest()[:8]
os.rename(filename, filename.replace(".pth", f"-{md5}.pth"))