---
license: cc-by-4.0
task_categories:
- image-segmentation
tags:
- remote-sensing
- earth-observation
- construction
- volume-estimation
pretty_name: EarthworkSat-7K
size_categories:
- 10K<n<100K
---

# EarthworkSat-7K

A satellite imagery dataset for finding earthwork on construction sites, and for estimating its volume.

About 7,400 annotated earthwork instances in 14 cities across Europe and Asia. The images are Google Earth screen captures at a pixel size of about 0.13 m, taken between 2018 and 2025. A separate set of 505 instances at 14 further locations is paired with public airborne LiDAR, so each of those instances has a reference volume in cubic metres.

![Annotated examples from the 14 cities](assets/fig_dataset.png)

## What is in the repository

```
benchmark/     24,596 image tiles (640 x 640 JPEG) with binary masks (PNG), in train / val / test, grouped by city code
annotations/   COCO polygon annotations for the same tiles, one JSON per split
volume/        415 image clips with masks, and one CSV per location with reference volumes
checkpoints/   six trained segmentation models (PyTorch)
```

Tiles and masks pair by filename: `images/CPH/src__CPH_2019_08_25_0001.jpg` goes with `masks/CPH/src__CPH_2019_08_25_0001_mask.png`. The split is by city: nine cities train, Oslo and Istanbul validate, and Copenhagen, Munich and Milan are the test set. The test cities are never seen in training, so the benchmark measures how a model transfers to a new city.

## Volume

For the 505 volume instances the mask marks the deposit and the CSV row gives its reference volume, computed from national airborne LiDAR by integrating height above the lowest point inside the boundary. Each CSV also has a `split` column: `finetune` if the instance was used to adapt the segmentation model to the volume imagery, `heldout` if it was kept aside for validation.

![The reference model on one full test capture](assets/fig_showcase.png)

## Benchmark results

Six segmentation models trained under the same recipe, evaluated on the three unseen test cities:

| Model | Foreground IoU | Dice |
|---|---|---|
| SegFormer-B5 | **0.747** | **0.855** |
| SegFormer-B3 | 0.736 | 0.848 |
| SegFormer-B4 | 0.735 | 0.847 |
| SegFormer-B2 | 0.728 | 0.842 |
| UPerNet (ConvNeXt-S) | 0.712 | 0.832 |
| DeepLabV3+ (ResNet-50) | 0.695 | 0.820 |

A volume pipeline (segment, align the mask to the elevation raster, integrate the height) reaches a median volume error of 5.6 percent on held-out instances.

## Checkpoints

`checkpoints/` holds the best checkpoint of each model above, saved as a PyTorch state dict:

| File | Size |
|---|---|
| `segformer_b5_msfe.pt` | 1.0 GB |
| `upernet_convnext.pt` | 0.9 GB |
| `segformer_b4.pt` | 0.7 GB |
| `segformer_b3.pt` | 0.5 GB |
| `deeplabv3plus_r50.pt` | 0.5 GB |
| `segformer_b2.pt` | 0.3 GB |

Each file contains `{"model_state_dict": ...}` for the corresponding architecture (SegFormer models follow the Hugging Face `transformers` implementation with a two-class head at 640 x 640 input).

## Scripts

Training and evaluation scripts, including the volume pipeline, will be uploaded here a little later.

## License and attribution

Annotations, reference volumes, benchmark splits and per-tile metadata are released under CC BY 4.0. Trained model weights and processing scripts are released under Apache 2.0.
The imagery tiles are screen captures from Google Earth; the underlying imagery is © Google and its providers and is redistributed here under CC BY-NC-ND 4.0 for non-commercial research use. The tiles are non-contiguous crops of individual earthwork sites and are not suitable for reconstructing a contiguous basemap. Users redistributing or building on the imagery should observe the providers' terms.Reference volumes derive from public national LiDAR programmes (AHN, UK Environment Agency, IGN, Ontario MNRF, Stadt Wien, and the Hong Kong 2020 LiDAR survey).

## Citation

The data descriptor is under review. A citation will appear here.