N
Selected work

Data & Research/2026/Prototype

Odum — Cocoa Deforestation Traceability

Helps cocoa exporters prove their beans are deforestation-free under the EU Deforestation Regulation. It reads Sentinel-2 satellite imagery to catch post-2020 forest loss at sub-hectare resolution, which is exactly where the usual 30 m datasets tend to miss smallholder farms, then produces a due-diligence statement with the evidence attached.

Context

Under the EU Deforestation Regulation, a cocoa exporter has to prove its beans are not tied to land cleared after December 2020. The catch is that the 30 m forest-loss datasets most tools lean on quietly miss the small farms that make up much of Ghana’s cocoa, so exporters are left without a defensible answer for regulators.

Approach

  • 01Built an offline machine-learning pipeline that ingests Sentinel-2 satellite imagery and detects post-2020 forest loss at sub-hectare resolution using a LightGBM change-detection model over weak labels.
  • 02Separated the system into an offline analysis stage and a read-only online serving layer (FastAPI) so heavy geospatial compute never blocks the dashboard.
  • 03Stored plots and verdicts in PostgreSQL/PostGIS and surfaced them through a Next.js + MapLibre console with per-district evidence tiles.
  • 04Generated a citation-anchored EUDR Due Diligence Statement as a signed PDF, tracing each verdict back to its imagery evidence.

Notable

  • Detects loss on plots that 30 m global datasets miss, closing the smallholder gap.
  • Multi-district console with per-plot verdicts and drill-down evidence tiles.
  • Regulator-ready, signed due-diligence output with source citations.

Next project

Sway — Influence Analytics Platform