README.md

---
license: cc-by-4.0
language:
- en
task_categories:
- image-to-image
- image-segmentation
pretty_name: GeoSR-Bench
size_categories:
- 10K<n<100K
---

# GeoSR-Bench

GeoSR-Bench directly connects super-resolution (SR) with downstream Earth monitoring tasks, moving beyond conventional fidelity-based evaluation. It comprises spatially co-located, temporally aligned, and quality-controlled image pairs from about 36,000 locations across diverse land covers, spanning spatial resolutions from 500 m to 0.6 m. It is designed to evaluate whether improved image resolution from SR models translates into better downstream performance for tasks such as land cover segmentation, infrastructure mapping, and biophysical variable estimation.
GeoSR-Bench includes two cross-platform super-resolution tasks:

- **MODIS → Landsat-8**
- **Sentinel-2 → NAIP**

For each task, the dataset is organized into two types of subsets:

1. **Super-resolution-only datasets**  
   These subsets include paired lower-resolution and higher-resolution remote sensing images without downstream task labels. They are designed for training SR models.

2. **Downstream task datasets**  
   These subsets include paired lower-resolution and higher-resolution images together with task-specific labels. They are designed to finetune SR models and evaluate whether super-resolved images improve downstream Earth monitoring tasks, such as land cover segmentation, infrastructure mapping, and biophysical variable estimation.

Each sample may contain:

- A lower-resolution image
- A higher-resolution reference image
- A downstream task label, when available
- Metadata, when available

GeoSR-Bench is intended to support research on task-aware super-resolution, cross-platform learning, and remote sensing foundation models.

## Folder Structure

The dataset contains both SR-only datasets and downstream task datasets.

### SR-only dataset

```text
SRDatasetName/
├── lr/ or modis/ or s2/
│   ├── lr_0.tif
│   ├── lr_1.tif
│   └── ...
├── hr/ or l8/ or naip/
│   ├── hr_0.tif
│   ├── hr_1.tif
│   └── ...
├── meta/
│   ├── meta_0.json
│   ├── meta_1.json
│   └── ...
└── SRDatasetName_split_all.csv
```

### Downstream task dataset

```text
DownstreamDatasetName/
├── s2/ or modis/
│   ├── s2_0.tif
│   ├── s2_1.tif
│   └── ...
├── naip/ or l8/
│   ├── naip_0.tif
│   ├── naip_1.tif
│   └── ...
├── label/
│   ├── label_0.tif
│   ├── label_1.tif
│   └── ...
├── meta/
│   ├── meta_0.json
│   ├── meta_1.json
│   └── ...
└── DownstreamDatasetName_split_all.csv
```

## Split Files

Each subset includes a CSV file describing the image paths and data split.

SR-only datasets are intended for training super-resolution models and do not have predefined train/validation/test split files.

For downstream task datasets, the CSV contains:

```text
LR image,HR image,Label,split
DownstreamDatasetName/s2/s2_0.tif,DownstreamDatasetName/naip/naip_0.tif,DownstreamDatasetName/label/label_0.tif,training
DownstreamDatasetName/s2/s2_1.tif,DownstreamDatasetName/naip/naip_1.tif,DownstreamDatasetName/label/label_1.tif,validation
DownstreamDatasetName/s2/s2_2.tif,DownstreamDatasetName/naip/naip_2.tif,DownstreamDatasetName/label/label_2.tif,test
```


## Data Fields

| Column | Description |
|---|---|
| `LR image` | Path to the lower-resolution input image |
| `HR image` | Path to the higher-resolution reference image |
| `Label` | Path to the downstream task label |
| `split` | Dataset split: `training`, `validation`, or `test` |

## File Formats

- Images are stored as GeoTIFF files (`.tif`).
- Labels are stored as GeoTIFF files (`.tif`).
- Metadata files, when available, are stored as JSON files (`.json`).
- Split files are stored as CSV files (`.csv`).

GeoTIFF files retain geospatial metadata such as coordinate reference system, transform, resolution, and spatial extent.

## Usage

You can read the split CSV using Python:

```python
import pandas as pd

csv_path = "DownstreamDatasetName/DownstreamDatasetName_split_all.csv"
df = pd.read_csv(csv_path)

train_df = df[df["split"] == "training"]
val_df = df[df["split"] == "validation"]
test_df = df[df["split"] == "test"]

print(len(train_df), len(val_df), len(test_df))
```

For SR-only datasets, use the `LR image` and `HR image` columns:

```python
sample = train_df.iloc[0]

lr_path = sample["LR image"]
hr_path = sample["HR image"]

print(lr_path)
print(hr_path)
```

For downstream task datasets, use the `LR image`, `HR image`, and `Label` columns:

```python
sample = train_df.iloc[0]

lr_path = sample["LR image"]
hr_path = sample["HR image"]
label_path = sample["Label"]

print(lr_path)
print(hr_path)
print(label_path)
```

You can read GeoTIFF files using `rasterio`:

```python
import rasterio

with rasterio.open("DownstreamDatasetName/s2/s2_0.tif") as src:
    image = src.read()
    crs = src.crs
    transform = src.transform

print(image.shape)
print(crs)
print(transform)
```

## Intended Use

This dataset is intended for research on:

- Remote sensing image super-resolution
- Downstream task-aware image restoration
- Land cover mapping
- Infrastructure mapping
- Biophysical variable estimation
- Cross-platform Earth observation learning
- Geo-foundation models

## Citation

If you use this dataset, please cite:

```bibtex
@article{li2026beyond,
  title={Beyond Visual Fidelity: Benchmarking Super-Resolution Models for Large-Scale Remote Sensing Imagery via Downstream Task Integration},
  author={Li, Zhili and Chai, Kangyang and Wang, Zhihao and Jia, Xiaowei and Li, Yanhua and Mai, Gengchen and Skakun, Sergii and Manocha, Dinesh and Xie, Yiqun},
  journal={arXiv preprint arXiv:2605.00310},
  year={2026}
}
```