Methodology
Where the data comes from, how detections are produced, and what they cannot tell you.
Data sources
- Sentinel-2 Level-2A
Copernicus surface reflectance, accessed via Google Earth Engine (COPERNICUS/S2_SR_HARMONIZED).
- Hansen Global Forest Change
Landsat-derived tree cover (2000) and annual loss year, University of Maryland / GLAD.
- ESA WorldCover
10 m land-cover map; used for context, never as the loss label.
Processing
- 1. Acquire For each tile of the chosen area, Earth Engine builds a median of cloud-masked Sentinel-2 scenes for the before season and the after season. Clouds, shadows, cirrus, saturation and snow are removed using the scene classification layer (SCL); every pixel keeps a count of clear observations.
- 2. Define forest Only pixels that were forest at the before date are scanned: canopy ≥ 30% in 2000 and no Hansen loss up to the before year.
- 3. Features Six reflectance bands per date, their differences, and NDVI, NBR and NDMI at both dates with their differences (27 features per pixel).
- 4. Three detectors A spectral-index rule (dNBR and dNDVI thresholds fitted on training sites), a Random Forest on the 27 features, and a Siamese U-Net whose weight-shared encoder compares before and after at every scale. Loss: weighted cross-entropy plus Dice, computed over forest pixels only.
- 5. Validate Whole sites are assigned to train, validation or test, so no test pixel neighbours a training pixel. Thresholds are chosen on validation sites. Precision, recall, F1, IoU, confusion matrices, a tile-bootstrap interval for F1 and calibration error are reported on test sites. The shown model is chosen by validation F1.
- 6. Vectorise Predicted loss pixels become 8-connected polygons; area is computed in a projected UTM grid at 10 m; each polygon gets a mean model probability and the dominant land-cover class.
Scientific limitations
- Hansen labels are a satellite-derived reference, not field truth. Reported accuracy is agreement with that product unless independent reference polygons were supplied.
- Optical imagery shows that canopy disappeared, not whether the clearing was legal. Any 'illegal' claim needs boundary and permit data that this system does not hold.
- Plantation harvest, fire and flooding can look like deforestation; WorldCover context helps triage but does not remove them.
- Selective logging and poaching of individual trees are below 10 m resolution and are not detected.
- Results hold for the evaluated sites, seasons and years. No worldwide generalisation is claimed.
Credits
Motion patterns (blur reveal, kinetic headline, line draw, node network) are adapted from the MIT-licensed NullMotion project by blixvip and contributors, re-implemented with Framer Motion.