Upload README.md with huggingface_hub
Browse files
README.md
CHANGED
|
@@ -15,8 +15,11 @@ tags:
|
|
| 15 |
- sentinel-2
|
| 16 |
- remote-sensing
|
| 17 |
- river
|
|
|
|
| 18 |
size_categories:
|
| 19 |
- 100K<n<1M
|
|
|
|
|
|
|
| 20 |
---
|
| 21 |
|
| 22 |
# EEA Waterbase Optics5 — Cleaned River Measurements (2015–2017)
|
|
@@ -42,6 +45,31 @@ Produced during the **11th CASSINI Hackathon – EU Space for Water**.
|
|
| 42 |
|
| 43 |
## Files
|
| 44 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 45 |
### `measurements_optics5_long.parquet`
|
| 46 |
Long-format table of all individual measurements. One row per (station, sampling date, determinand).
|
| 47 |
|
|
@@ -96,9 +124,10 @@ Audit table: number of measurements per station, year, and determinand code.
|
|
| 96 |
|
| 97 |
## Intended use
|
| 98 |
|
| 99 |
-
Each training sample for a Sentinel-2 model is a `(station, date)` pair joined to the nearest cloud-free Sentinel-2 scene within a ±N day window. Separate models (or a multitask model with target masking) are trained per determinand — co-presence of all five codes on the same date is not required.
|
| 100 |
|
| 101 |
## Source
|
| 102 |
|
| 103 |
-
Derived from `cassini-team-todo/eea-waterbase` (mirror of EEA Waterbase WISE-4 v2018.1).
|
|
|
|
| 104 |
Original data © European Environment Agency, reused under the [EEA reuse policy](https://www.eea.europa.eu/en/legal-notice).
|
|
|
|
| 15 |
- sentinel-2
|
| 16 |
- remote-sensing
|
| 17 |
- river
|
| 18 |
+
- geoparquet
|
| 19 |
size_categories:
|
| 20 |
- 100K<n<1M
|
| 21 |
+
modality:
|
| 22 |
+
- geospatial
|
| 23 |
---
|
| 24 |
|
| 25 |
# EEA Waterbase Optics5 — Cleaned River Measurements (2015–2017)
|
|
|
|
| 45 |
|
| 46 |
## Files
|
| 47 |
|
| 48 |
+
### `station_river_bboxes.geoparquet`
|
| 49 |
+
**Primary spatial file for model training.** One row per station matched to its nearest HydroRIVERS segment. The geometry is a ~5 km corridor polygon that follows the river shape — suitable as a Sentinel-2 query window with the station at the midpoint.
|
| 50 |
+
|
| 51 |
+
Construction:
|
| 52 |
+
1. Each station is snapped onto its nearest HydroRIVERS segment (max 5 km)
|
| 53 |
+
2. All connected river geometry within 2.5 km of the snap point is merged into a continuous linestring
|
| 54 |
+
3. A 5 km sub-segment centred on the snap point is extracted (2.5 km each direction)
|
| 55 |
+
4. The segment is buffered 300 m each side → corridor polygon in WGS-84
|
| 56 |
+
|
| 57 |
+
| Column | Description |
|
| 58 |
+
|---|---|
|
| 59 |
+
| `station_id` | Monitoring site identifier |
|
| 60 |
+
| `lat`, `lon` | Original station coordinates (WGS84) |
|
| 61 |
+
| `first_date` | Earliest measurement date in window |
|
| 62 |
+
| `last_date` | Latest measurement date in window |
|
| 63 |
+
| `n_measurements` | Total number of measurements |
|
| 64 |
+
| `n_determinands` | Number of distinct Optics5 codes measured |
|
| 65 |
+
| `HYRIV_ID` | Matched HydroRIVERS segment ID |
|
| 66 |
+
| `MAIN_RIV` | HydroRIVERS main-river ID |
|
| 67 |
+
| `snap_dist_m` | Distance from station to the river line (metres) |
|
| 68 |
+
| `geometry` | Corridor polygon (Polygon, WGS84) following the river shape |
|
| 69 |
+
| `bbox_west/south/east/north` | Envelope coordinates of the corridor for Sentinel Hub API calls |
|
| 70 |
+
|
| 71 |
+
**4,742 stations** (15 dropped — farther than 5 km from any river)
|
| 72 |
+
|
| 73 |
### `measurements_optics5_long.parquet`
|
| 74 |
Long-format table of all individual measurements. One row per (station, sampling date, determinand).
|
| 75 |
|
|
|
|
| 124 |
|
| 125 |
## Intended use
|
| 126 |
|
| 127 |
+
Each training sample for a Sentinel-2 model is a `(station, date)` pair joined to the nearest cloud-free Sentinel-2 scene within a ±N day window. The corridor polygon in `station_river_bboxes.geoparquet` defines the spatial extent of the Sentinel-2 query — pass `bbox_west/south/east/north` directly to the Sentinel Hub Statistical API. Separate models (or a multitask model with target masking) are trained per determinand — co-presence of all five codes on the same date is not required.
|
| 128 |
|
| 129 |
## Source
|
| 130 |
|
| 131 |
+
Derived from `cassini-team-todo/eea-waterbase` (mirror of EEA Waterbase WISE-4 v2018.1).
|
| 132 |
+
River network from [`cassini-team-todo/hydro-rivers-europe`](https://huggingface.co/datasets/cassini-team-todo/hydro-rivers-europe) (HydroRIVERS v1.0).
|
| 133 |
Original data © European Environment Agency, reused under the [EEA reuse policy](https://www.eea.europa.eu/en/legal-notice).
|