GeoNDC

GeoNDC uses ML and satellite data to track NO₂, forest loss, and gas flaring, delivering transparent scores, rankings, and alerts so African governments target NDC actions fast.

Comment

Inspiration:

  • Africa’s NDC progress is hard to track: scattered data, slow reports, little clarity on what drives change.
  • We wanted a fast, transparent signal that shows where emissions are rising or falling this month, not last year.

What it does:

  • Monitors three satellite-observable drivers: NO2 (near-real-time combustion proxy), gas flaring, and forest loss.
  • Uses ML to separate weather effects from human activity in NO2.
  • Scores and ranks countries/regions, flags hotspots, shows trend arrows and confidence.
  • Exports a transparent report with data sources, formulas, and assumptions.

How we built it:

  • Data: Sentinel-5P/TROPOMI for NO2, VIIRS Nightfire for flaring, Global Forest Watch/GLAD for forest loss, ERA5/NOAA reanalysis for weather.
  • Model: LightGBM predicts expected NO2 from weather and seasonality; anomalies = observed minus expected.
  • Score: weighted blend of NO2 anomaly, flaring trend, and forest-loss rate; weights reflect evidence for sector contributions.
  • Confidence: satellite coverage and cloudiness, agreement with ground stations where available, and cross-validation.
  • Stack: Python, LightGBM, scikit-learn, pandas/xarray, PostGIS, FastAPI backend, React dashboard.

Challenges we ran into:

  • Satellite gaps and clouds causing patchy coverage.
  • Class imbalance for rare NO2 spikes.
  • Harmonizing spatial and temporal resolutions across sensors.
  • Communicating uncertainty simply to decision-makers.

Accomplishments we’re proud of:

  • A working pipeline producing monthly anomaly maps, country scores, and confidence bands.
  • Clear methodology docs and an ethics/QA checklist (data provenance, versioning, model cards).
  • Threshold tuning that improved recall on real emission events without flooding users with false alarms.

What we learned:

  • NO2 gives actionable, near-real-time signals, but context from weather, season, and policy is essential.
  • Transparency builds trust; users adopt scores they can audit.
  • Small UX choices (default thresholds, copy) shape behavior more than fancy models.

What’s next for GeoNDC:

  • Add methane (TROPOMI CH4) and power-plant activity layers; expand to more countries.
  • Calibrate with additional ground monitors and partner ministries for pilots.
  • Launch an alerts API, richer “what-if” policy scenarios, and third-party methodology review.
  • Formalize ethics features: bias monitoring, data-use governance, and per-score confidence explanations.

Built With

Share this project:

Updates