| Source | Grain | Status | Use | Reason for classification |
|---|---|---|---|---|
| INPE PRODES deforestation | 6.25 ha, annual, 1988-present | Usable | Long-term deforestation baseline (reference) | Official Brazilian government deforestation monitoring system using Landsat-class imagery. Gold-standard reference for Amazon deforestation. 6.25 ha minimum mapping unit misses small clearings. Cloud cover may delay detection in wet season. Brazilian Amazon only; other Amazon countries have no PRODES equivalent. |
| Hansen et al. global forest change | 30 m, annual, 2000-present | Usable | Global tree cover loss and gain baseline | First global 30 m forest change dataset from Landsat time series (Science 2013, DOI: 10.1126/science.1244693). Annual updates via Global Forest Watch. Measures tree cover loss, not deforestation; plantation harvest and fire counted as loss. Forest definition threshold (30 percent canopy) affects results. |
| GLAD forest alerts (Hansen/U Maryland) | 30 m, weekly, 2019-present | Usable | Near-real-time new forest loss detection | Weekly Landsat-based alerts distributed via Global Forest Watch. Enables rapid response monitoring. Detects tree cover loss generically — does not distinguish deforestation, fire, or selective logging. 30 m resolution may miss understory degradation. |
| Global Forest Watch platform | Variable, weekly to annual | Usable | Integrated forest monitoring data access | WRI platform integrating Hansen loss maps, GLAD alerts, GLAD-S2 alerts, and fire alerts with downloadable data and APIs. Aggregates products with different definitions, resolutions, and update frequencies that must not be conflated. Does not provide land cover classification. |
| GBIF species occurrence | Point, variable | Usable with caveats | Species occurrence for range estimation | Over 2.7 billion global occurrence records, millions in Amazon basin. Open-access with APIs. Occurrence density is heavily skewed toward research stations and rivers; vast interior regions are severely under-sampled. Records reflect sampling effort, not species distribution. Many historical records have imprecise coordinates. |
| IUCN Red List range maps | Polygon, species-level | Usable with caveats | Species range for conservation status | Expert-drawn range maps for 46,000+ assessed species including ~2,300 Amazon vertebrates. Range maps are extent-of-occurrence polygons that overestimate actual occupancy. Static snapshots may not reflect recent range contraction. Not all Amazon species assessed; biased toward vertebrates. Intersection with 30 m deforestation data implies false precision at range edges. |
| MapBiomas land cover | 30 m, annual, 1985-present | Usable | Land cover change classification | Brazilian land cover and land use maps from Landsat ML classification. 25+ classes with deforestation-to-pasture and deforestation-to-agriculture transition matrix. Brazil-only; does not cover Peruvian, Colombian, Ecuadorian, or Bolivian Amazon. Classification accuracy varies by biome and year. ML can produce temporal inconsistencies. |
| Terra-i / RADD radar alerts | 10 m (Sentinel-1), weekly | Limited | Radar-based deforestation alerts (cloud-penetration) | Terra-i and RADD provide radar-based deforestation alerts that penetrate cloud cover, addressing the key limitation of optical systems. However, radar alert methodology is newer with less validation history than GLAD. Classification accuracy in Amazon floodplain and wetland areas is uncertain. Not yet integrated as a reference dataset. |
| MODIS Vegetation Continuous Fields | 250 m, annual, 2000-present | Rejected (grain) | Coarse vegetation change context | 250 m resolution is too coarse for Amazon deforestation detection where small clearings under 6.25 ha dominate. Useful for broad regional trend context, not for pixel-level deforestation mapping or species-range intersection. Rejected for canonical modeling; listed as context. |
- Date range.
- Geographic grain.
- Deforestation detection method including minimum mapping unit.
- Species occurrence data including collection date and location precision.
- License or reuse permission.
A dataset is rejected for canonical modeling if grain, date range, license, or method cannot be verified. Rejected datasets may still be listed as context.
The critical data gap for Amazon deforestation-species risk analysis is the mismatch between deforestation detection at 30 m resolution and species range maps at coarse polygon scale. GBIF occurrence data is point-based but spatially biased and temporally inconsistent. No public dataset provides repeated population-level survey data for Amazon species that would allow direct estimation of deforestation-driven population decline. Range-loss estimates built from IUCN polygons intersected with Hansen or PRODES deforestation will produce results with large uncertainty bands that must be explicitly communicated, not hidden behind precise-looking maps.