ToDo

TODO — Satellite-Based Pollution Triage for European Rivers

💎 Idea

Europe has 160,576 industrial facilities discharging to water and no continental-scale system to tell which downstream communities and ecosystems are actually affected. We built a platform that links every facility in the EEA's E-PRTR database to its surrounding river network and uses Copernicus Sentinel-2 to detect water quality changes upstream and downstream of each site.

The core idea is to treat each facility as its own controlled experiment. For every facility within 1 km of a mapped river, the system extracts two geographic chunks — 10 km upstream and 10 km downstream of the discharge point, traced along the actual flow-direction graph. Upstream is the control. Downstream is the signal. When downstream water diverges from its upstream baseline in ways consistent with pollution (turbidity, chlorophyll, water extent), the facility is flagged for inspection.

This isolates each facility's local impact from basin-wide noise like seasonal change or weather, which is what makes it pollution detection rather than general water monitoring.

🛰️ EU space technologies

Our pipeline is built on three datasets:

• Sentinel-2 L2A via the Copernicus Data Space Ecosystem — accessed through the Sentinel Hub Statistical API on 10-day aggregation at 10 m resolution. A custom evalscript computes NDWI, NDCI (chlorophyll-a proxy), and a turbidity ratio per pixel, masked to water surfaces (NDWI > 0). Bands used: B02, B03, B04, B05, B08, B8A.
• EU-Hydro River Network Database from the Copernicus Land Monitoring Service — used to determine which river segments are wide enough to monitor at Sentinel-2's 10 m resolution.
• HydroRIVERS v1.0 (WWF/USGS) — provides the river line network with flow direction for upstream/downstream tracing on the NEXT_DOWN graph.

ML models trained on EEA Waterbase in-situ measurements correlate band ratios with specific water quality determinands.

🌊 EU Space for Water

Primary challenge: Tracking and preventing water pollution.

Our contribution:

• Source-to-impact traceability — first open dataset linking every E-PRTR facility to its upstream and downstream river geometry.
• Per-facility upstream-as-control design — each facility's downstream signal is compared against its own upstream baseline, isolating local impact from basin-wide noise.
• Continental-scale screening — 160,576 facilities is too many to inspect manually. The system provides data-driven triage so regulators inspect the highest-risk cases first.
• Sentinel-2's 5-day revisit catches episodic discharges that monthly in-situ sampling misses.
• Browser-native deployment — the 2.6 GB GeoParquet dataset is queried directly in the browser via DuckDB WASM, with no backend server. Marginal cost per user is near zero.

Secondary fit: equitable water access — downstream municipalities, irrigators, and ecosystems benefit from visible, attributable pollution impacts.

VIDEO DEMO HERE (https://youtu.be/e7wvjnyC0Oo)

🤼 Team

Four-person team, all M.Sc. Data Science at TU Wien:

• Tomáš Hobza — Geospatial pipeline lead. B.Sc. FIT VUT Brno; previously Microsoft, Gen Digital, FNZ; Lead Software Developer at Metafide; incoming Cloudflare intern (Lisbon).
• Jakub Vsetecka — Sentinel statistics, ML training, web app. B.Sc. FIT VUT Brno; previously ML engineer at CESNET (DDoS packet filtering); intern at BAWAG.
• Martina Adamuščinová — Data exploration and Google Earth Engine experiments. B.Sc. Economics from WU Vienna (Data Science, Business Information Systems, Climate Economics); Research Project Assistant at BONARD.
• Kai Thilenius — Sentinel fetch scripts. B.Eng. Electrical Engineering from DHBW Mosbach (3-year co-op at Bosch Rexroth); hardware testing intern at Bosch Austria.