README.md

metadata

license: cc-by-4.0
task_categories:
  - tabular-classification
  - feature-extraction
tags:
  - environment
  - rivers
  - pollution
  - geospatial
  - europe
size_categories:
  - 100K<n<1M

EEA Industrial Emissions - River Proximity Dataset

This dataset links 160,576 industrial facilities from the European Environment Agency (EEA) to nearby rivers, with upstream/downstream flow analysis based on HydroRIVERS network data.

Why This Dataset?

Industrial facilities that discharge pollutants into water bodies affect downstream ecosystems and communities. This dataset enables:

• Impact assessment: Which communities/ecosystems are downstream of polluting facilities?
• Source tracing: Where does a facility's water supply come from?
• Regulatory analysis: Mapping industrial emissions to affected river networks
• Environmental research: Studying relationships between industry and water quality

How It Works

         ┌─────────────┐
         │  UPSTREAM   │  ← Source water (green)
         │  (5 parts)  │     flowing TOWARD facility
         └──────┬──────┘
                │
         ───────●─────── ← Split point (closest to facility)
                │
         ┌──────┴──────┐
         │  FACILITY   │  ← Industrial facility (orange)
         │   (211m)    │     within 1km of river
         └─────────────┘
                │
         ┌──────┴──────┐
         │ DOWNSTREAM  │  → Affected water (red)
         │  (4 parts)  │     flowing AWAY from facility
         └─────────────┘

For each facility:

1. Find the closest river segment within 1km
2. Split the river at the closest point
3. Trace upstream 10km (source water)
4. Trace downstream 10km (potentially affected)
5. Clip water surface polygons to match

Dataset Files

river_data_facilities.geoparquet (2.6 GB)

Main dataset with 160,576 facilities matched to rivers.

Column	Type	Description
facilityName	string	Name of industrial facility
city	string	City location
countryName	string	Country
EPRTR_SectorCode	int	Industry sector code (1-9)
EPRTR_SectorName	string	Industry sector name
Pollutant	string	Pollutant released to water
Releases	float	Amount released
closest_river_id	int	HydroRIVERS segment ID
distance_to_river_m	float	Distance to river (meters)
river_strahler	int	Strahler stream order (1-9)
river_discharge	float	Average discharge (m³/s)
upstream_segment_ids	list[int]	Upstream HydroRIVERS IDs
downstream_segment_ids	list[int]	Downstream HydroRIVERS IDs
n_upstream	int	Number of upstream parts
n_downstream	int	Number of downstream parts
upstream_line_wkb	bytes	Upstream river geometry (WKB)
downstream_line_wkb	bytes	Downstream river geometry (WKB)
upstream_poly_wkb	bytes	Upstream water surface (WKB)
downstream_poly_wkb	bytes	Downstream water surface (WKB)
geometry	Point	Facility location (WGS84)
closest_river_overlap_fraction	float	Fraction (0–1) of the closest HydroRIVERS segment's length that falls inside an EU-Hydro River_Net_p polygon (± 30 m buffer)
closest_river_is_sentinel_visible	bool	True if closest_river_overlap_fraction ≥ 0.30 — the closest river segment is wide enough to be detectable in Sentinel-2
n_upstream_sentinel_visible	int	Number of upstream segments that meet the Sentinel-visibility threshold
n_downstream_sentinel_visible	int	Number of downstream segments that meet the Sentinel-visibility threshold
has_sentinel_visible_river	bool	Summary flag — True if any of the closest, upstream, or downstream segments is Sentinel-visible; use this column to filter facilities to those where Sentinel-2 water-quality retrieval is feasible

river_data_segments.geoparquet (1.7 MB)

River segments with direction labels (28,434 entries).

Column	Type	Description
HYRIV_ID	int	HydroRIVERS segment ID
direction	string	"upstream" or "downstream"
ORD_STRA	int	Strahler stream order
DIS_AV_CMS	float	Average discharge (m³/s)
LENGTH_KM	float	Segment length (km)
geometry	LineString	River segment geometry

Usage

Load the dataset

import geopandas as gpd
from shapely import wkb

# Load facilities
facilities = gpd.read_parquet("river_data_facilities.geoparquet")
print(f"Loaded {len(facilities):,} facilities")

# Example: Find facilities in Germany
german = facilities[facilities['countryName'] == 'Germany']
print(f"Germany has {len(german):,} facilities near rivers")

Extract river geometries

# Get a specific facility
facility = facilities[facilities['facilityName'].str.contains('PRECHEZA')].iloc[0]

# Parse WKB geometries
upstream_line = wkb.loads(facility['upstream_line_wkb'])
downstream_line = wkb.loads(facility['downstream_line_wkb'])
upstream_poly = wkb.loads(facility['upstream_poly_wkb'])
downstream_poly = wkb.loads(facility['downstream_poly_wkb'])

print(f"Upstream: {facility['n_upstream']} segments")
print(f"Downstream: {facility['n_downstream']} segments")

Visualize a facility

python visualize_single_facility.py "PRECHEZA"
# Opens facility_map.html in browser

Visualize multiple facilities

python visualize_facilities_rivers.py
# Opens facilities_rivers_map.html with 200 sampled facilities

Scripts

river_proximity.py

Main pipeline that:

1. Loads EEA facilities, HydroRIVERS segments, and EU-Hydro polygons
2. Builds river network graph from NEXT_DOWN field
3. For each facility, finds closest river and splits at nearest point
4. Traces upstream (BFS) and downstream (linear) within distance limits
5. Clips water surface polygons to match river geometries
6. Outputs geoparquet files

visualize_single_facility.py

Creates an interactive Folium map for a single facility showing:

• Green: upstream river and water surface
• Red: downstream river and water surface
• Orange marker: facility location

visualize_facilities_rivers.py

Creates an overview map with sampled facilities and their river associations.

Source Data

Dataset	Source	Usage
Industrial Facilities	EEA E-PRTR	Facility locations & emissions
River Network	HydroRIVERS v1.0	River segments & flow direction
Water Polygons	EU-Hydro	Water surface geometry

Parameters

Parameter	Value	Description
max_distance_m	1,000	Max distance from facility to river
upstream_distance_km	10	How far to trace upstream
downstream_distance_km	10	How far to trace downstream
polygon_buffer_m	600	Buffer for polygon clipping

Statistics

• Total facilities processed: 254,027
• Facilities near rivers: 160,576 (63%)
• Unique upstream segments: 18,498
• Unique downstream segments: 9,936
• Average upstream parts: 7.1
• Average downstream parts: 3.7

Sentinel visibility (EU-Hydro River_Net_p overlap ≥ 30 %)

Strahler order	Facilities	% Sentinel-visible
1	43,699	22.5 %
2	28,113	20.5 %
3	31,616	23.9 %
4	24,961	42.5 %
5	15,709	67.8 %
6	8,128	86.5 %
7	8,131	97.5 %
8	219	100.0 %

• Facilities with has_sentinel_visible_river = True: 59,568 / 160,576 (37.1 %)
• Facilities where the closest reach is visible: 23,507 / 160,576 (14.6 %)

Lower-order rates (~20–25 %) are driven by upstream/downstream propagation: a headwater facility may drain into a wider river within the 10 km trace window.

facility_timeseries.parquet (210 MB)

Sentinel-2 water-quality time series for Sentinel-visible industrial facilities (both upstream and downstream polygons detectable in 10 m imagery). Produced by fetching the Sentinel Hub Statistical API in P10D bins over 2017–2023 and merging three shards. One row per (facility, direction, 10-day bin).

Column	Type	Description
facility_id	int	Row index in river_data_facilities.geoparquet
direction	string	"upstream" or "downstream"
polygon_hash	string	SHA-256[:16] of the raw WKB polygon bytes (dedup key)
date	datetime	Start of the 10-day bin
ndci_mean	float	Mean NDCI over water pixels in bin
ndci_stddev	float	Std dev of NDCI
ndci_n_valid	int	Number of valid (water-masked) pixels
turb_mean	float	Mean turbidity proxy over water pixels
turb_stddev	float	Std dev of turbidity
turb_n_valid	int	Number of valid pixels (turbidity)
ndwi_mean	float	Mean NDWI (water index, quality check)
ndwi_n_valid	int	Number of valid pixels (NDWI)

Coverage: 9.2 M rows · 23,158 unique facility IDs · 20,755 paired (both directions) · 2017-01-01 → 2023-12-16

facility_anomalies_per_bin.parquet (216 MB)

Per-bin anomaly detection output. One row per paired (facility, 10-day bin) where both upstream and downstream data exist. Produced by scripts/compute_facility_anomalies.py.

Pipeline steps applied:

1. Inner-join upstream + downstream on (facility_id, date)
2. Pixel-count quality flags
3. Raw downstream-minus-upstream delta
4. Spatial detrending: subtract cross-facility median delta per date to remove regional Sentinel-2 artifacts
5. Robust z-score per (facility_id, quarter) using median + MAD × 1.4826
6. Previous-bin z (persistence check)
7. High-confidence anomaly flag (detrended delta > 0, z > 3, prev-z > 1.5)

Column	Type	Description
facility_id	int	Facility identifier
date	datetime	10-day bin start
quarter	int	Calendar quarter (1–4), used for seasonal baseline
poly_hash_upstream	string	Upstream polygon hash
poly_hash_downstream	string	Downstream polygon hash
ndci_mean_upstream / _downstream	float	NDCI means per direction
ndci_n_valid_upstream / _downstream	int	Valid pixel counts
turb_mean_upstream / _downstream	float	Turbidity means
turb_n_valid_upstream / _downstream	int	Valid pixel counts
valid_bin_ndci / valid_bin_turb	bool	Passes pixel-count quality filter
delta_ndci_raw	float	Raw downstream − upstream NDCI
date_median_ndci	float	Cross-facility median NDCI delta on this date (removed artifact)
delta_ndci	float	Spatially detrended NDCI delta
delta_turb_raw / delta_turb	float	Same for turbidity
baseline_med_ndci / baseline_mad_ndci	float	Seasonal baseline median and MAD
z_delta_ndci / z_delta_turb	float	Robust z-scores of detrended deltas
z_delta_ndci_prev / z_delta_turb_prev	float	Previous-bin z (persistence)
low_baseline_data	bool	True if < 8 bins in the seasonal group
high_confidence_ndci / high_confidence_turb	bool	Anomaly flag per signal
any_anomaly	bool	Either signal flagged
event_key	string	Links to event in facility_anomalies_events.parquet (null if not in a kept event)

Coverage: 4.26 M rows · 20,755 facilities · 255 unique dates

facility_anomalies_events.parquet (tiny)

Consolidated pollution events — one row per unique (upstream polygon, downstream polygon, time window) after deduplication. Single-bin events and events where the downstream signal is not worse than upstream are excluded.

Column	Type	Description
event_id	string	Unique event key (e.g. 12345_e3)
facility_id	int	Representative facility for this polygon pair
start_date	datetime	First flagged bin
end_date	datetime	Last flagged bin
duration_bins	int	Number of flagged 10-day bins
peak_z_ndci	float	Maximum z-score (NDCI) across bins in event
peak_z_turb	float	Maximum z-score (turbidity) across bins in event
mean_z_ndci / mean_z_turb	float	Mean z-scores over event
signal_type	string	"ndci", "turb", or "both"
poly_hash_upstream / poly_hash_downstream	string	Physical polygon pair (dedup key)

Coverage: 42 events · 36 unique polygon pairs · 2017–2023

Thresholds used: z > 3.0, prev-bin z > 1.5, ≥ 2 consecutive bins, detrended delta > 0, seasonal MAD baseline requires ≥ 8 bins per quarter.

License

CC-BY-4.0. See source datasets for their respective licenses.

Citation

If you use this dataset, please cite the source datasets:

• Lehner, B., Grill G. (2013): Global river hydrography and network routing: baseline data and new approaches to study the world's large river systems. Hydrological Processes, 27(15): 2171–2186.
• European Environment Agency (EEA) Industrial Emissions Database
• Copernicus EU-Hydro River Network Database