Deforestation is one of the few environmental problems you can see from space, consistently and globally. That's why regulations like the EUDR lean on satellite data, and why compliance has become partly a remote sensing problem.
The basic workflow
For compliance, the question is narrow: was this plot forest on the cut-off date, and has it been deforested since?
- Forest baseline: was it forest on 31 December 2020?
- Change detection: has forest been lost since?
- Plot overlay: does the loss intersect the plot?
- Review: does a human agree?
Data sources
| Dataset | Type | Notes |
|---|---|---|
| JRC Global Forest Cover 2020 | Baseline | Published by the EU's Joint Research Centre in support of EUDR |
| Global Forest Change (Hansen et al., UMD) | Annual tree cover loss | Landsat-based, from 2000 |
| JRC Tropical Moist Forests | Deforestation and degradation | Tropics |
| GLAD alerts | Near-real-time loss | Landsat / Sentinel-2 based |
| RADD alerts | Near-real-time disturbance | Sentinel-1 radar, works through clouds |
| Global Forest Watch | Platform | Aggregates many of the above |
Plot overlay in practice
With plots stored in PostGIS and a loss raster converted to polygons (or sampled via a raster tool):
SELECT p.id, SUM(ST_Area(ST_Intersection(p.geom, l.geom)::geography)) / 10000 AS loss_ha FROM plots p JOIN forest_loss_after_2020 l ON ST_Intersects(p.geom, l.geom) GROUP BY p.id;
Then classify:
- No overlap: low risk (still check the baseline said "forest" or "not forest" sensibly).
- Small overlap at edges: often geometry error or mixed pixels, so review.
- Substantial overlap: high risk, needs investigation and evidence.
For point-only plots, buffer the point to an approximate area, but treat results as weaker evidence.
Where it goes wrong
- Resolution: 30 m pixels on a 1-hectare plot means about 11 pixels. Edges matter a lot.
- Clouds: optical alerts are delayed in cloudy seasons. Radar helps.
- Definitions: rubber, cocoa agroforestry, and plantations confuse forest classifiers.
- Harvest vs deforestation: timber harvest followed by regrowth can look like loss.
- Geometry errors: a misplaced polygon creates a false positive. Validate first (how).
Human review
Build a review queue:
- Show the plot over before and after imagery.
- Show which dataset flagged it and when.
- Let reviewers attach evidence (field visits, documents, higher-resolution images).
- Record the decision and reason.
This is both good practice and an audit trail.