README.md

---
license: other
license_name: bsd-4-clause
license_link: LICENSE
language:
  - en
task_categories:
  - object-detection
tags:
  - table-detection
  - table-structure-recognition
  - historical-documents
  - weather-records
  - scanned-documents
size_categories:
  - n<1K
configs:
  - config_name: coarse
    data_files:
      - split: train
        path: coarse/train-*.parquet
      - split: test
        path: coarse/test-*.parquet
  - config_name: fine
    data_files:
      - split: train
        path: fine/train-*.parquet
      - split: test
        path: fine/test-*.parquet
---

# GloSAT Table Dataset

Table structure recognition annotations for scanned historical meteorological logbook pages from the [GloSAT project](https://www.glosat.org/). Images are scanned pages from 9 archival sources spanning the 19th and early 20th centuries. Each image is annotated with table bounding boxes and per-cell bounding boxes with row/column indices and header flags.

This dataset supports table detection and table structure recognition (TSR) tasks on historical document images.

**No transcribed cell content is included.** This is a structure-only dataset: it provides spatial layout (bounding boxes and grid topology) but not the text values inside cells.

## Configs

| Config | Description | Cells per table (mean) | Cells per table (max) |
|---|---|---|---|
| `coarse` | Merged cells treated as single units | 156 | 476 |
| `fine` | Each physical cell annotated individually | 488 | 2737 |

The two configs cover the same 499 images with the same train/test split. Use `coarse` for tasks that treat spanning cells as atomic units; use `fine` for tasks that require individual cell boundaries.

## Splits

| Split | Images |
|---|---|
| train | 371 |
| test | 129 |

## Schema

Each row represents one scanned page image.

| Field | Type | Description |
|---|---|---|
| `image_id` | string | Numeric image identifier (matches original Zenodo filenames) |
| `image` | Image | Embedded JPEG bytes |
| `width` | int32 | Image width in pixels |
| `height` | int32 | Image height in pixels |
| `source` | string | Source archive name (see provenance table below) |
| `tables` | Sequence | List of annotated tables on the page |
| `tables.bbox` | Sequence[int32] | Table bounding box: `[xmin, ymin, xmax, ymax]` in image pixel coordinates |
| `tables.type` | string | Table border style (e.g., `semi-bordered`) |
| `tables.cells` | Sequence | List of annotated cells within the table |
| `tables.cells.bbox` | Sequence[int32] | Cell bounding box: `[xmin, ymin, xmax, ymax]` |
| `tables.cells.row_start` | int32 | Zero-indexed start row |
| `tables.cells.row_end` | int32 | Zero-indexed end row (exclusive) |
| `tables.cells.col_start` | int32 | Zero-indexed start column |
| `tables.cells.col_end` | int32 | Zero-indexed end column (exclusive) |
| `tables.cells.header` | bool | True if the cell is a header cell |

All coordinates are in original image pixel space. 39% of train images contain more than one table per page.

## Source provenance

| Source | Images | Description |
|---|---|---|
| `DWR` | 93 | Daily Weather Reports |
| `Ben_Nevis` | 97 | Ben Nevis Observatory records |
| `WR_10_years` | 97 | 10-year weather registers |
| `WesTech_Rodgers` | 82 | WesTech / Rodgers logbooks |
| `WR_Devon_Extern` | 33 | Devon external weather registers |
| `20cr_Natal_Witnes` | 26 | 20th Century Reanalysis: Natal Witness |
| `20cr_DWR_MO` | 24 | 20th Century Reanalysis: DWR (Met Office) |
| `20cr_DWR_NOAA` | 24 | 20th Century Reanalysis: DWR (NOAA) |
| `20cr_Kubota` | 24 | 20th Century Reanalysis: Kubota |

## Usage

```python
from datasets import load_dataset

# Fine-grained cell annotations
ds = load_dataset("rootsautomation/GloSAT", "fine")
row = ds["train"][0]

print(row["image_id"], row["source"], row["width"], row["height"])
# Access table annotations
for i, bbox in enumerate(row["tables"]["bbox"]):
    cells = row["tables"]["cells"][i]
    print(f"Table {i}: bbox={bbox}, {len(cells['bbox'])} cells")
```

## License

This dataset is released under a BSD 4-Clause license (see `LICENSE`). The 4-Clause variant includes an **advertising clause**: any materials or publications that use or reference this dataset must acknowledge the GloSAT project in their advertising or documentation.

Original data source: [Zenodo record 5363457](https://zenodo.org/records/5363457).

## Citation

If you use this dataset, please cite the original GloSAT paper:

```bibtex
@inproceedings{glosat2021,
  title     = {GloSAT Historical Measurement Table Dataset: End-to-End Table Processing with Columns Delimited at Run-Time},
  author    = {Middleton, Stuart E. and Kordopatis-Zilos, Giorgos and Sheridan, Iain},
  booktitle = {Proceedings of the 4th International Workshop on Historical Document Imaging and Processing (HIP)},
  year      = {2021},
  doi       = {10.5281/zenodo.5363457}
}
```