Orion-AI-Lab / KuroSiwo

Code and data for Kuro Siwo flood mapping dataset

Kuro Siwo: A global multi-temporal SAR dataset for rapid flood mapping

Latest updates:

- [✔️] More events outside of Europe (43 in total)
- [✔️] We included the respective SLC products and cropped patches in Kuro Siwo
- [✔️] Downloading script and links have been updated for the new version
- [✔️] Preprocessing pipelines for both GRD and SLC data can be found in `configs/`
- [✔️] Updated paper: https://arxiv.org/abs/2311.12056
- [ ] TODO: minor updates to training and dataloading code 

Table of Contents

• Download the dataset
• Data preprocessing
• Repository structure
• Pretrained models
• Citation

Download Kuro Siwo

GRD Data

• The Kuro Siwo GRD Dataset can be downloaded either:

  • from the following link,

  • or by executing scripts/download_kuro_siwo.sh. This script will download and prepare the Kuro Siwo GRDD dataset for deep learning.

    Usage

    1. Make sure to grant the necessary rights by executing chmod +x scripts/download_kuro_siwo.sh
    2. Execute scripts/download_kuro_siwo.sh DESIRED_DATASET_ROOT_PATH e.g: ./download_kuro_siwo.sh KuroRoot

SLC Data

• The SLC Preprocessed products can be downloaded from the following link.

• Similarly, the cropped SLC patches (224x224 pixels) can be acquired from the following link.

Data preprocessing

The preprocessing pipelines used to generate the GRD and SLC products can be found at configs/grd_preprocessing.xml and configs/slc_preprocessing.xml repsectively.

Kuro Siwo repo structure

• Kuro Siwo uses the black python formatter. To activate it install pre-commit, running pip install pre-commit and execute pre-commit install.
• Training starts by running python main.py. The configurations are defined in the configs directory e.g
  • model,
  • training pipeline
    • Segmentation,
    • change detection
  • hyperparameters
• main.py supports command line arguments that override the config files. e.g

     python main.py --method=unet --backbone=resnet18 --dem=True --slope=False --batch_size=32
  

Pretrained models

The weights of the top performing models can be accessed using the following links:

• FloodViT
• SNUNet

Citation

If you use this work please cite:

@inproceedings{Bountos2023KuroS3,
  title={Kuro Siwo: 33 billion \$m^2\$ under the water. A global multi-temporal satellite dataset for rapid flood mapping},
  author={Nikolaos Ioannis Bountos and Maria Sdraka and Angelos Zavras and Ilektra Karasante and Andreas Karavias and Themistocles Herekakis and Angeliki Thanasou and Dimitrios Michail and Ioannis Papoutsis},
  year={2023},
  url={https://api.semanticscholar.org/CorpusID:265308874}
}