If you find our work or benchmark helpful for your research, please consider giving us a ⭐!
- 🛎️ Updates
- ✨ Requirements
- 🛠️ Installation
- 📄 Datasets
- 📄 Method
- ✨ Checkpoints
- 📄 Results
- 📚 Citation
- 🔑 License
Before using this repository, make sure you have the following prerequisites installed:
torch==2.6.0 torchaudio==2.6.0 torchvision==0.21.0
Tested using Python 3.9.0 on a 64-bit Windows operating system.
conda install pytorch torchvision pytorch-cuda=12.4 -c pytorch -c nvidiaTo get started, create the conda environment named SAFE-FD by executing the following command:
conda create -n SAFE_FD python=3.9.0 -yThen activate the environment:
conda activate SAFE_FDDownload the datasets and place them in the Benchmark folder.
| Dataset | Name | Link |
|---|---|---|
| Ombria-Net | OmbriaS1 |
source |
| DAM-Net | S1GFloods |
source |
| New-Dataset | S1GFloods+7 |
baidu drive |
Prepare the following folders to organize this repo:
SAR_Flood_Detection
├── SAFENet (code)
├── PRETRAINED (pvt_v2_b2.pth)
├── WORK_DIR (save the model weights and training logs)
│ └─ checkpoints (the best models)
│ └─ test_metrics (the best test results)
│ ├─test_metrics.txt (the best test results)
│ └─ test_visualizations
│ └─ train_history (train & val results per epoch)
│ └─ val_results
└── Benchmark
├── OmbriaS1
│ ├── train
│ │ ├── A Images of Time 1 before the flood event
│ │ │ └── S1_0001.png
│ │ ├── B Images of Time 2 after the flood event
│ │ │ └── S1_0001.png
│ │ └── GT Ground truth labels
│ │ └── S1_0001.png
│ ├── test (the same with train)
│
├── SIGFloods
│ ├── train
│ │ ├── A
│ │ │ └── S1_1.png
│ │ ├── B
│ │ │ └── S1_1.png
│ │ └── GT
│ │ └── S1_1.png
│ ├── test (the same with train)
│
├── SIGFloods+7
│ ├── train
│ │ ├── A
│ │ │ └── Brazil_0_1.png
│ │ │ └── Indonesia_0_1.png
│ │ │ └── Morocco_0_1.png
│ │ │ └── Mozambique_0_1.png
│ │ │ └── Portugal_0_1.png
│ │ │ └── Ukraine_0_1.png
│ │ │ └── USA_0_1.png
│ │ ├── B
│ │ │ └── Brazil_0_1.png
│ │ │ └── Indonesia_0_1.png
│ │ │ └── Morocco_0_1.png
│ │ │ └── Mozambique_0_1.png
│ │ │ └── Portugal_0_1.png
│ │ │ └── Ukraine_0_1.png
│ │ │ └── USA_0_1.png
│ │ └── GT
│ │ └── Brazil_0_1.png
│ │ │ └── Indonesia_0_1.png
│ │ │ └── Morocco_0_1.png
│ │ │ └── Mozambique_0_1.png
│ │ │ └── Portugal_0_1.png
│ │ │ └── Ukraine_0_1.png
│ │ │ └── USA_0_1.png
│ ├── test (the same with train)
| Method | Name | Link |
|---|---|---|
| 📖 📖 📖 DBF-Net | DBF-Net |
Paper here |
| 📖 📖 📖 WBA-Net | WBA-Net |
Paper here |
| 📖 📖 📖 GLA-Former | GLA-Former |
Paper here |
| 📖 📖 📖 CAS-Net | CAS-Net |
Paper here |
| 📖 📖 📖 PA-Former | PA-Former |
Paper here |
| 📖 📖 📖 ViCxLSTM | ViCxLSTM |
Paper here |
| 📖 📖 📖 BiSR-Net | BiSR-Net |
Paper here |
| 📖 📖 📖 SCan-Net | SCan-Net |
Paper here |
| Method | Input size | F1-score | Checkpoints |
|---|---|---|---|
| DBF-Net | 256 × 256 | 53.3 | Model |
| WBA-Net | 256 × 256 | 52.9 | Model |
| GLA-Former | 256 × 256 | 65.7 | Model |
| CAS-Net | 256 × 256 | 71.3 | Model |
| PA-Former | 256 × 256 | 76.7 | Model |
| ViCxLSTM | 256 × 256 | 77.4 | Model |
| BiSR-Net | 256 × 256 | 78.0 | Model |
| SCan-Net | 256 × 256 | 81.3 | Model |
| SAFE-Net | 256 × 256 | 85.3 | Model |
| Method | Input size | F1-score | Checkpoints |
|---|---|---|---|
| DBF-Net | 256 × 256 | 62.4 | Model |
| WBA-Net | 256 × 256 | 88.7 | Model |
| GLA-Former | 256 × 256 | 93.6 | Model |
| CAS-Net | 256 × 256 | 86.5 | Model |
| PA-Former | 256 × 256 | 95.9 | Model |
| ViCxLSTM | 256 × 256 | 95.0 | Model |
| BiSR-Net | 256 × 256 | 94.2 | Model |
| SCan-Net | 256 × 256 | 96.2 | Model |
| SAFE-Net | 256 × 256 | 96.8 | Model |
-
OmbriaS1
Qualitative comparison on the OmbriaS1 dataset. From left to right: T1 image, T2 image, Ground Truth, DBF-Net, CAS-Net, WBA-Net, GLA-Former, BiSR-Net, ViCxLSTM, PA-Former, SCan-Net, and the proposed SAFE-Net method (Ours).
-
S1GFloods
Qualitative comparison on the S1GFloods dataset. From left to right: T1 image, T2 image, Ground Truth, DBF-Net, CAS-Net, WBA-Net, GLA-Former, BiSR-Net, ViCxLSTM, PA-Former, SCan-Net, and the proposed SAFE-Net method (Ours).
-
S1GFloods+7
@ARTICLE{tasaleh2026SAFENet,
Author = {Tamer Saleh, Gui-Song Xia, Wen Yang, Mohamed Ihmeida, Zhuohong Li, Shimaa Holail},
Title = {SAFE-Net: A Shifted-Attention Frequency-Gated Network and Large-Scale Benchmark for Generalizable SAR Flood Mapping},
Journal = {ISPRS Journal of Photogrammetry and Remote Sensing},
Year = {2026},
volume={},
number={},
pages={1-24},
doi = {},
}
@article{saleh2024dam,
title={DAM-Net: Flood detection from SAR imagery using differential attention metric-based vision transformers},
author={Saleh, Tamer and Weng, Xingxing and Holail, Shimaa and Hao, Chen and Xia, Gui-Song},
journal={ISPRS Journal of Photogrammetry and Remote Sensing},
volume={212},
pages={440--453},
year={2024},
publisher={Elsevier}
}If you are confused about the content of our paper or look forward to further academic exchanges and cooperation, please do not hesitate to contact us. The e-mail address is tamersaleh@whu.edu.cn. We look forward to hearing from you!
licensed for research use only. The dataset is CC BY NC 4.0 (allowing only non-commercial use), and models trained using the dataset should not be used outside of research purposes.