PEN03 - Show Me the Beach. I'll Show You the Source.

MARÉ is a forensic intelligence system for marine litter that answers three questions no coastal municipality can currently answer: where did the litter on the beach come from, when will the next wave of debris arrive, and where to act upstream to prevent it. The solution combines inverse Lagrangian particle modelling over open oceanographic data from Copernicus Marine with AI-based visual classification of debris and independent biological validation through biofouling, triangulating ocean physics, computer vision and marine biology. For each beach, it returns a report with probable source zones, contribution percentages and 48-72h forecasts. The timing is favourable: Portugal spends an estimated €10-20M/year on reactive coastal cleanup with no source attribution tools, and the second cycle of the EU Marine Strategy Framework Directive (2027) makes source identification mandatory for Portugal's 30 coastal municipalities. The business model is annual B2G subscription (€8-50k/client: municipality, region, or national agency), targeting €1M in annual recurring revenue within 24 months in Portugal alone, with expansion to Spain and the EU Atlantic arc (~€20M). The team combines a marine biologist (Miguel Santos) and an AI analyst (Nuno Santos), a pairing that makes the methodology scientifically defensible and the product operational from day one. 100% open-data stack: Copernicus, OceanParcels, ERA5, TACO. No lock-in, no commercial dependencies.

Over the 48 hours we will build a working proof-of-concept of the MARÉ platform — an interactive web prototype backed by a real data-processing pipeline that demonstrates the full forensic attribution workflow for one anchor case (Praia da Consolação, Peniche, January 2026 storm scenario).

Concretely, we will develop four components:
1. A Lagrangian backtracking algorithm in Python, using OceanParcels and the Copernicus Marine Iberia hydrodynamic model, that simulates 5,000 inverse particles from the target beach over a 7-day window, identifies probable source zones, and outputs attribution percentages per zone.
2. An AI-based debris classifier built on a pre-trained YOLOv8 model and the open TACO dataset, capable of categorising photographed beach litter into source-relevant typologies (domestic/fluvial, fishing gear, industrial, hygiene), which then weight the Lagrangian attribution.
3. An interactive web dashboard (Streamlit) that fuses the two outputs: for the anchor beach, it displays the source attribution map, contribution percentages by origin zone, debris composition charts, and an actionable recommendation panel for the municipal customer.
4. A reproducible demonstration scenario — the January 2026 storm at Consolação — with synthetic field data calibrated against public reports from APA, Quercus and Sciaena, used to walk the jury through the end-to-end forensic flow during the final pitch.

The output is not a production SaaS but a credible end-to-end demonstrator that proves the methodology works, the data sources integrate, and the user-facing report is intelligible to a non-technical municipal customer.

marine biologist, AI analyst