Noise2Map

Official implementation of Noise2Map: End-to-End Diffusion Model for Semantic Segmentation and Change Detection (IEEE TGRS 2026).

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

[Paper] · [Pretrained Weights]

---

Overview

Noise2Map repurposes the diffusion denoising trajectory for discriminative remote sensing tasks. Rather than iterative sampling, it predicts semantic or change maps in a single forward pass using task-specific structured noise schedules.

• Semantic Segmentation (SS): single image → segmentation mask
• Change Detection (CD): bi-temporal image pair → change mask

Key results (ranked 1st on both tasks across SpaceNet7, WHU, and xView2-Wildfire):

Task	SpaceNet7 F1	WHU F1	xView2 F1
SS	72.02	95.69	86.90
CD	71.43	95.27	86.91

---

Installation

git clone https://github.com/alishibli97/noise2map.git
cd noise2map
pip install -e .

Or install dependencies directly:

pip install -r requirements.txt

---

Pretrained Weights

The denoising UNet backbone is pretrained on 10k images from the AID dataset using a DDPM denoising objective. Weights are hosted on HuggingFace and loaded automatically at runtime.

Config key	HuggingFace path	Description
aid_10k	ali97/noise2map / aid-10k	AID pretrained, min-max norm, Google DDPM init (recommended)
satellite	ali97/noise2map / sat2gen	Pretrained on MajorTOM satellite imagery
imagenet	ali97/noise2map / imagenet2gen	ImageNet pretrained
google	ali97/noise2map / ddpm-church	Google DDPM church checkpoint
None	—	Random initialisation (ablation)

See HuggingFace for all pretrained weights.

---

Datasets

Download datasets and place them under data/:

data/
├── whu-cd/          # WHU Building CD  — train/{A,B,OUT}/  test/{A,B,OUT}/
├── WHU/             # WHU Building SS  — train/{Image,Mask}/  test/{Image,Mask}/
├── xview2/          # xView2 Wildfire  — train/xBD/santa-rosa-wildfire/{images,masks}/
└── spacenet7/       # SpaceNet7        — train/<aoi_id>/{images_masked,labels_raster}/

Dataset-specific filtering (e.g. WHU-CD skips image pairs with no changes) is handled automatically and cached on first run.

---

Training

Each experiment is configured via a YAML file in configs/. To train:

python train.py --config configs/whu_cd.yaml
python train.py --config configs/whu_sem.yaml
python train.py --config configs/xview2_cd.yaml
python train.py --config configs/xview2_sem.yaml
python train.py --config configs/spacenet7_cd.yaml
python train.py --config configs/spacenet7_sem.yaml

Override any config value on the command line:

python train.py --config configs/whu_cd.yaml training.batch_size=4 training.num_epochs=100

SLURM example scripts are in scripts/.

---

Evaluation

python evaluate.py --config configs/whu_cd.yaml --checkpoint path/to/best_model.pth

Outputs per-class and mean F1, IoU, Precision, Recall, and Accuracy.

---

Citation

@article{shibli2026noise2map,
  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},
}

---

Acknowledgement

This research is part of the EO-AI4GlobalChange project funded by Digital Futures, Stockholm, Sweden.