Noise2Map — Pretrained Backbones

Pretrained denoising UNet backbones for Noise2Map: End-to-End Diffusion Model for Semantic Segmentation and Change Detection (IEEE TGRS 2026).

Ali Shibli, Andrea Nascetti, Yifang Ban — KTH Royal Institute of Technology

[GitHub]

---

Checkpoints

Subfolder	Description
aid-10k	Pretrained on 10k AID aerial images (recommended)
sat2gen	Pretrained on MajorTOM Sentinel-2 satellite imagery
imagenet2gen	ImageNet pretrained
ddpm-church	Google DDPM church-256

---

Usage

from noise2map import Noise2Map

model = Noise2Map(
    in_channels=6,   # 3 for semantic segmentation
    out_channels=2,
    img_scale=256,
    pretrained="aid_google_minmaxnorm",
)

See the GitHub repo for full training and evaluation instructions.

---

Citation

@article{shibli2025noise2map,
  title   = {Noise2Map: End-to-End Diffusion Model for Semantic Segmentation and Change Detection},
  author  = {Shibli, Ali and Nascetti, Andrea and Ban, Yifang},
  journal = {IEEE Transactions on Geoscience and Remote Sensing},
  year    = {2026},
}