README.md

---
license: cc-by-4.0
pretty_name: BloomField
tags:
- remote-sensing
- water-quality
- chlorophyll-a
- harmful-algal-blooms
- spatial-reconstruction
- benchmark
size_categories:
- 1B<n<10B
---

# BloomField

Whole-lake chlorophyll-*a* (Chl-*a*) **field-reconstruction** benchmark for the
**trusted-sparse × biased-dense** regime: a handful of accurate in-situ stations paired with two
independent, value-biased ocean-colour satellites (GOCI-II geostationary + Sentinel-3 OLCI), over
three eutrophic lakes in eastern China (Tai, Chao, Hongze), 2021–2024.

## Structure
Per-lake folder (`tai/`, `chao/`, `hongze/`):
- `goci_cache.npy`, `olci_cache.npy` — gridded per-scene Chl-*a* rasters on a common ~300 m lake grid.
- `goci_scenes.json`, `olci_dates.json` — scene → cache-row indices.
- `goci/`, `sentinel/` — raw per-scene rasters and lake-grid metadata (`lake_meta*`).
- `situ/` — in-situ station time series (chlorophyll-*a*, turbidity, and station coordinates only).
- `split_index.csv`, `alignment_index.csv` — (scene, station) match-ups with temporal `split`.
- `goci_station_features.csv`, `olci_station_features.csv` — station-pixel satellite values.
- `station_meta.csv`, `shp/` — station and lake geometry.

Splits are strictly temporal (train ≤2023, val 2024-H1, test 2024-H2).

## Usage
The benchmark code consumes a lake via two environment variables:
```bash
export SCENE_DATA_ROOT=/path/to/bloomfield/tai
export SCENE_META=/path/to/bloomfield/tai/goci/full/lake_meta_taihu.npz
```
Download:
```python
from huggingface_hub import snapshot_download
snapshot_download(repo_id="ALH84001/bloomfield", repo_type="dataset", local_dir="bloomfield")
```

## Companion resources
The benchmark **code** and the small **station × satellite match-up tables** are released separately;
their links are provided upon publication (anonymized during double-blind review).

## License & sources
Released under **CC-BY-4.0**. Underlying Level-2 products (GOCI-II, Sentinel-3 OLCI) are public and
re-downloadable; in-situ readings come from an automatic-analyser station network. See the paper's
Dataset Datasheet appendix for full provenance, collection, and ethics.

## Citation
> BloomField: a benchmark for whole-lake chlorophyll-*a* field reconstruction. *Under review.*

## Raw per-scene rasters (tarballs)
The `*_cache.npy` files already stack every scene for direct use by the code, so most users need only
the caches. For provenance / custom re-gridding we also ship the individual per-scene `.npz` rasters as
per-lake tarballs (Lake Tai GOCI scenes are cache-only):
- `<lake>/scenes_sentinel.tar`, `<lake>/scenes_goci.tar`

Extract inside the lake folder to restore `sentinel/scene_npz/` and `goci/full/scene_npz/`:
```bash
cd tai && tar -xf scenes_sentinel.tar
```