A geospatial information system that combines free satellite data and in-situ monitoring to map the phytoplankton bloom risk along the Romagna coast (northern Adriatic), with quantified uncertainty on every prediction, and a causal analysis of the effect of the Po river discharge on chlorophyll.

Technical report: a working-paper writeup of this project (method, results, honest positioning and limitations) is available at docs/technical-report.md, and as a citable preprint on Zenodo (DOI above).

Interactive web map: model-estimated bloom risk and real Copernicus Marine chlorophyll-a per coastal cell, with the detail panel showing the prediction, its 90% interval and the cell threshold.

flowchart TD
    A["Public data sources<br/>Copernicus Marine · ERA5 · GloFAS"] --> B["Feature pipeline<br/>xarray, geopandas"]
    B --> C["Predictive model<br/>LightGBM + conformal prediction<br/>(point estimate + confidence interval, always together)"]
    C --> D["PostGIS<br/>coastal cells, stations, predictions with native uncertainty"]
    D --> E["FastAPI<br/>/api/risk · /api/stations · /api/chlorophyll → GeoJSON"]
    E --> F["Leaflet web map<br/>colour = risk, hatching = uncertainty"]
    B --> G["Causal layer (causal/)<br/>A: transparent estimate · B: fixed effects<br/>C: DoWhy + refuters · D: causal forest"]

• Docker and Docker Compose
• Python >= 3.11
• A free Copernicus Marine account (chlorophyll, SST)
• A free ECMWF account (ERA5 wind on the CDS, GloFAS Po discharge on the EWDS)

The pipeline downloads data from two services, each with its own credentials.

• Copernicus Marine: run copernicusmarine login once (interactive).
• Climate Data Store / Early Warning Data Store: put your ECMWF token in ~/.cdsapirc:

  url: https://cds.climate.copernicus.eu/api
  key: <YOUR-TOKEN>
  

  The same token works for both stores; the EWDS URL is set inside the script. Each dataset licence must be accepted once on its download page.

make install

Creates a virtualenv in .venv and installs the pipeline dependencies.

make run

• Map: http://localhost:8000
• API: http://localhost:8000/api/risk

The system starts with demo data (5 cells along the coast, from the Po mouth to Cattolica) - see db/init.sql. No real data is needed to see it working.

With the stack running:

make ingest      # download all public data, multi-year (long: CDS/EWDS queue)
make features    # build the feature table
make train       # train the model, write predictions, replace the demo data
make run         # rebuild the API to serve the new predictions

make causal      # Step A (transparent) + Step B (DoWhy with refuters)

See causal/README.md for the DAG, results and assumptions.

Satellite data (Sentinel, Copernicus, ERA5) and Po discharge are public and are not stored in the repository (they are reproduced by the scripts in pipeline/). In-situ phytoplankton data from the ARPAE-Daphne monitoring network, if used, is requested under the Italian environmental-information act (D.Lgs. 195/2005) and is not redistributed here pending licence and terms of use. Anyone can request it from the ARPAE-Daphne oceanographic unit.

• api/ - FastAPI service (serves GeoJSON + the static web map)
• db/ - PostGIS schema and demo seed data
• pipeline/ - ingestion, feature engineering, training
• webmap/ - Leaflet web map (static front-end)
• causal/ - causal analysis layer
• tests/ - API tests
• .github/ - CI (lint + tests)

• Skeleton: PostGIS, API, web map, demo data
• Layer 1: real chlorophyll ingestion (Copernicus Marine) + map layer
• Layer 2: robust cleaning + drivers (SST, wind, Po discharge) + features
• Layer 3: LightGBM + conformal, derived risk, real predictions on the map
• Multi-year scaling: ingestion and features over several seasons, tested on an unseen year
• Layer 7 (causal/): Po effect estimated (Step A transparent + Step B DoWhy with refuters)
• Step C (causal/): effect heterogeneity via causal forest (EconML), spatial pattern found (effect concentrated near the Po delta), no interpretable temporal trend

Planned web-map polishing, to complete in the refinement phase:

• Show the date of the displayed data on the map.
• Clarify in the legend that the cells and grid are a demonstrative simplification (not the real transects).
• Make the header more informative (data period/coverage).

This project does not claim a novel ML method: gradient boosting, conformal prediction and - in the next layer - causal ML are all established methods, already applied to algal blooms in the literature. The contribution here is fine geographic scale (the Romagna transects, not the whole Adriatic), operational integration (a public, browsable system, not just a study), and rigour on uncertainty (every prediction carries a validated interval).

Declared limits: remote sensing estimates chlorophyll/biomass, not species; the northern Adriatic coastal waters are optically complex (Case-2) and need local calibration; observational causal estimates rely on assumptions (no unobserved confounding) that are declared and discussed, not taken for granted.

Code under the MIT license (see LICENSE). Environmental data remains subject to its original licenses.