README.md

---
license: cc-by-nc-sa-4.0
viewer: false
task_categories:
- image-segmentation
tags:
- river_ice
size_categories:
- 100K<n<1M
---

# NWPU YRCC GFICE

<!-- Provide a quick summary of the dataset. -->

**NWPU_YRCC_GFICE (GFICE)** is a satellite remote-sensing semantic segmentation dataset for **fine-grained river-ice mapping**. It is built from **multi-spectral GF-2 satellite imagery** over the **Yellow River (Ningxia–Inner Mongolia section)** and is designed to support hydrological monitoring and ice-flood early warning research.

## Dataset Details

### Dataset Description

<!-- Provide a longer summary of what this dataset is. -->

The NWPU_YRCC_GFICE dataset provides **pixel-level annotations** for river-ice scenes captured from satellite view, addressing the limited spatial coverage of prior river-ice datasets that mainly rely on UAV or ground-based imagery. The dataset covers **freeze–thaw cycles** and annotates river-ice conditions into **eight fine-grained semantic classes**.

- **Curated by:** Northwestern Polytechnical University (NWPU) research team (authors of the companion paper)
- **License:** CC BY-NC-SA 4.0 (`cc-by-nc-sa-4.0`)

## Uses

<!-- Address questions around how the dataset is intended to be used. -->

### Direct Use

<!-- This section describes suitable use cases for the dataset. -->

- Training and evaluating **semantic segmentation** models for **river ice mapping** from satellite imagery.
- Research on **fine-grained ice characterization** across **freeze–thaw cycles**.
- Developing algorithms for **hydrological monitoring** and **ice-flood early warning** (research/non-commercial use under the dataset license).
- Benchmarking segmentation architectures (the companion work evaluates many segmentation models and includes improved YOLO-/SegFormer-based variants).

### Out-of-Scope Use

<!-- This section addresses misuse, malicious use, and uses that the dataset will not work well for. -->

- **Commercial use** without meeting the constraints of **CC BY-NC-SA 4.0**.
- Direct deployment to other rivers/regions/sensors without considering **domain shift** (different illumination, sediment, width, climate, sensor bands, spatial resolution, etc.).

## Dataset Structure

<!-- This section provides a description of the dataset fields, and additional information about the dataset structure such as criteria used to create the splits, relationships between data points, etc. -->

The dataset is intended for **semantic segmentation** with **mmsegmentation** dataset style. A typical sample contains:

- `image`: a multi-spectral GF-2 TIFF image
- `label`: a per-pixel label map with **8 classes** (integer IDs)

**Splits (recommended):**
- `train` set contains 22700 images
- `validation` set contains 5675 images

**Label schema:**
The dataset contains **eight fine-grained classes** for river ice across freeze–thaw cycles.

| class_id | class_name |
|---:|---|
| 0 | land |
| 1 | water |
| 2 | shore_ice |
| 3 | other_ice |
| 4 | snow_covered_ice |
| 5 | stream_ice |
| 6 | ice_in_river |
| 7 | ice_outside_river |

## Citation

<!-- If there is a paper or blog post introducing the dataset, the APA and Bibtex information for that should go in this section. -->

**BibTeX:**
```bibtex
@ARTICLE{11145880,
  author={Wei, Chenxu and Li, Haoxuan and Chen, Liang and Zhou, Haohao and Taukebayev, Omirzhan and Wu, Wencong and Temirbayev, Amirkhan and Han, Lin and Ran, Lingyan and Yin, Hanlin and Wang, Peng and Liu, Junrui and Zhang, Xiuwei and Zhang, Yanning},
  journal={IEEE Transactions on Geoscience and Remote Sensing}, 
  title={River Ice Fine-Grained Segmentation: A GF-2 Satellite Image Dataset and Deep Learning Benchmark}, 
  year={2025},
  volume={63},
  number={},
  pages={1-15},
  keywords={Ice;Rivers;Satellites;Monitoring;Feature extraction;Transformers;Semantic segmentation;Accuracy;Benchmark testing;Remote sensing;Fine-grained semantic segmentation;river ice dataset;SegFormer;YOLO},
  doi={10.1109/TGRS.2025.3604644}
}
```

> Addition Note: We found that a few images in the dataset associated with the paper contained processing errors, and we have corrected them in the currently released public version of the dataset.