Annual tree cover loss and gain tracking

Annual global tree cover change is now measurable at 30 m resolution from 2000 onwards, but the numbers mean different things to different jurisdictions. Definitional gaps between remote-sensing products and national forest inventories have direct consequences for REDD+ carbon accounting.

Sensors

Landsat 8/9 OLI: 30 m spatial resolution, 16-day repeat at the equator (8-day when both satellites are combined). The backbone of the Hansen/UMD Global Forest Change product. Near-infrared and shortwave-infrared bands distinguish photosynthetically active canopy from bare soil and regrowth. Free archive from 1972 (MSS) and continuous from 1999 (ETM+) provides the longest consistent time-series available.
Sentinel-2 MSI: 10 m resolution in visible and near-infrared bands, 5-day revisit at mid-latitudes when both Sentinel-2A and 2B are operational. Finer spatial detail resolves small-scale disturbances and narrow riparian strips that fall below Landsat's detection floor. Free and open under Copernicus; increasingly used to validate and sharpen Hansen-derived loss polygons.
MODIS Terra/Aqua: 250 m to 500 m resolution, near-daily global coverage. Too coarse for 30 m loss mapping but useful for compositing cloud-free reference periods and for cross-checking annual loss totals at regional scale. The MODIS Vegetation Continuous Fields product provides independent canopy-cover fraction estimates.
Planet SuperDove: 3 m resolution, near-daily revisit over most of the tropics. Not used in the Hansen baseline product but increasingly applied to verify loss polygons, detect sub-hectare clearings, and reduce cloud-gap uncertainty in persistently overcast regions. Commercial licence required.

What the Hansen product actually measures, and what it does not

The University of Maryland Global Forest Change dataset, first published in Science in 2013 and updated annually, maps tree cover loss and gain globally at 30 m resolution from the year 2000 onwards using dense Landsat time-series stacks. The methodology classifies each 30 m pixel against a baseline canopy-cover percentage derived from the year-2000 epoch, then detects subsequent loss (removal of canopy) and gain (establishment of canopy where none existed in 2000) using spectral change detection and, in later versions, gradient-boosted machine-learning classifiers trained on manually interpreted samples.

The product defines tree cover as any woody vegetation taller than five metres, regardless of land use. A eucalyptus plantation, a shaded coffee farm, and an old-growth rainforest all contribute equally to the canopy-cover layer at the same percentage. Loss of a plantation rotation therefore appears identical in the data to clearance of primary forest unless a separate land-use or forest-type mask is applied. This is not a flaw in the methodology; it is an honest representation of what spectral reflectance can and cannot distinguish at 30 m.

The 10% threshold problem and what it means for REDD+

The Hansen product reports canopy cover in percentage bins and allows users to set their own threshold for what counts as 'forested'. The most widely cited global figures use a 10% canopy-cover threshold. Brazil's official national forest definition uses 30% canopy cover with a minimum area of 1 hectare and a minimum height of 5 metres. The Food and Agriculture Organisation's Forest Resource Assessment uses 10% canopy cover but also requires a minimum area of 0.5 hectares and potential to reach 5 m height. These differences are not academic: applying the 10% Hansen threshold to Brazilian Amazon data produces a larger reported forest area than the national definition, which in turn affects baseline calculations for REDD+ credits.

Countries reporting under Article 5 of the Paris Agreement must reconcile satellite-derived products with their own nationally determined forest definitions. A government that purchases Hansen-derived loss statistics without adjusting for its own threshold will overcount or undercount deforestation relative to its official inventory, creating a discrepancy that auditors and carbon registries will flag. The definitional gap is manageable, but it requires explicit masking and documentation at the analysis stage.

Cloud contamination in the tropics: the accuracy ceiling nobody advertises

Landsat's 16-day revisit sounds adequate until you consider that the Congo Basin, the Indonesian archipelago, and the western Amazon can be cloud-covered for more than 200 days per year. A single cloud-free observation per year is sometimes all that a Landsat time-series contains for a given pixel. The Hansen methodology handles this by building annual composites from all available clear observations, but in persistently cloudy regions the composite may rest on only one or two usable scenes. Loss events that occur and are subsequently re-vegetated within a cloudy gap can be missed entirely.

Published accuracy assessments of the Hansen product report overall accuracies above 90% at the global scale, but tropical forest loss accuracy in high-cloud regions is substantially lower and varies by year depending on the number of cloud-free observations available. Sentinel-2's shorter revisit partially addresses this, but cloud persistence in the wet tropics affects all optical sensors equally. Synthetic aperture radar, covered in a sibling page on sub-canopy damage mapping, is the only orbital technology that sees through cloud reliably, though it introduces its own classification ambiguities.

Annual versus near-real-time: choosing the right product for the question

Annual products like Hansen are designed for consistent long-term accounting, not rapid response. The annual loss layer is released with a lag of several months after the reference year closes, which is appropriate for national inventory submissions and trend analysis but unsuitable for enforcement or early warning. Near-real-time alert systems such as GLAD Alerts (also from UMD, using Landsat and Sentinel-2) and RADD Alerts (Sentinel-1 SAR) operate at latencies of days to weeks and are covered in a separate page in this library.

For annual accounting purposes the Hansen product remains the standard reference. Its continuous archive from 2000 gives analysts a 24-year baseline, long enough to separate cyclical disturbance from structural deforestation trends and to attribute loss to drivers such as agricultural expansion, fire, or selective logging when combined with auxiliary layers.

Gain is harder than loss, and the numbers show it

Tree cover gain in the Hansen product is mapped only for the full 2000-to-present epoch, not year by year. Detecting gain requires that a pixel cross the canopy-cover threshold from below, which takes years of regrowth and demands multiple consistent observations to confirm. Annual gain mapping at 30 m remains an open research problem: spectral signatures of young regrowth overlap with agricultural crops, particularly in the first two to four years after disturbance, and a single anomalous observation can generate a false positive.

This asymmetry matters for net-deforestation reporting. Gross loss figures from Hansen are frequently cited in policy documents; net figures that subtract gain are less common and less reliable. A country with active plantation expansion alongside primary forest clearance will look very different depending on whether net or gross accounting is applied. Analysts working on REDD+ baselines should treat gain figures with more caution than loss figures and cross-validate against national inventory data where available.

Applying the data: what a structured analysis pipeline looks like

A credible annual tree-cover change analysis for a national client involves several steps beyond downloading the Hansen tiles. The raw loss layer must be intersected with a primary forest mask (such as the UMD primary humid tropical forest layer), a land-use classification, and the client country's official forest definition threshold. Loss attributed to fire is separated from loss attributed to clearing using auxiliary burned-area products. Gain pixels are validated against Sentinel-2 time-series to filter false positives from seasonal crops.

Satellize runs this kind of structured pipeline on open constellations, applying the same approach it uses in its Tonga crop-estimation programme: combining multiple open-data sources with a clearly documented methodology so that the output is auditable by a third party. The deliverable is not a raw raster but an interpreted, jurisdiction-specific change report with uncertainty bounds attached to every headline figure. Buyers presenting deforestation statistics to international bodies need that audit trail; the satellite data alone does not provide it.

Typical figures

Spatial resolution (primary product)	30 m (Landsat-based Hansen/UMD); 10 m with Sentinel-2 validation layer
Revisit interval	16 days per satellite (Landsat 8 or 9); ~8 days combined; 5 days with Sentinel-2A+2B
Annual product latency	Hansen annual update typically released 3-6 months after reference year close
Archive depth	Landsat continuous from 1972 (MSS), systematic global coverage from 1999 (ETM+); Hansen loss layers from 2000
Spectral bands used	Near-infrared, shortwave-infrared (SWIR 1 and 2), red, green (OLI bands 3-7); NDVI and NBR indices derived
Canopy-cover threshold (Hansen default)	10% (user-adjustable; national definitions commonly use 25-30%)
Minimum detectable loss patch	Approximately 0.09 ha (single 30 m pixel); practical detection floor closer to 0.5-1 ha in cloud-affected regions
Reported global accuracy (loss layer)	Above 90% overall at global scale; lower in persistently cloudy tropical regions (figures vary by year and region in published assessments)
Coverage	Global land surface; Hansen product covers approximately 143 million km² of land
Delivery formats	GeoTIFF tiles (Hansen public release); jurisdiction-specific GeoPackage, Shapefile or PostGIS layer for processed outputs

Analytics Satellize can run

Annual gross tree-cover loss by jurisdiction	Hansen/UMD loss layer intersected with administrative boundaries and client-specified canopy-cover threshold; fire-attributed loss separated using MODIS burned-area product	Annual change report (PDF + GeoPackage) with uncertainty bounds and threshold-sensitivity table
Primary forest loss versus plantation loss disaggregation	Hansen loss layer masked against UMD primary humid tropical forest layer and a land-use classification derived from Sentinel-2 composites	Stratified area statistics GIS layer with per-stratum confidence intervals
Canopy-cover fraction time-series (2000 to present)	Annual Landsat composites processed through vegetation continuous fields model; NDVI and SWIR2 band ratios calibrated against MODIS VCF reference	Per-pixel time-series raster stack and summary trend chart by administrative unit
Net change accounting (loss minus gain)	Hansen gain layer (epoch-level) apportioned across years using Sentinel-2 regrowth phenology; false-positive filtering via multi-date spectral trajectory	Net forest balance table with documented methodology note for submission to national inventory or carbon registry
REDD+ baseline forest definition reconciliation	Threshold sensitivity analysis: Hansen loss and gain areas recalculated at 10%, 25% and 30% canopy-cover thresholds and compared against national forest inventory reference figures	Reconciliation memo and adjusted area statistics for use in nationally determined contribution reporting
Multi-decadal deforestation driver attribution	Spatial overlay of loss polygons with agricultural expansion layers, road buffers, fire perimeters and protected-area boundaries; gradient-boosted classifier following published driver-attribution literature	Driver-attributed loss map and summary statistics by driver class and time period

Sources

Who does the work

We can get this done for you. Satellize runs its own analyst desk and a strong science team. You do not buy a data feed and work out what it means; our people source the imagery, run the analysis described on this page, and hand you the answer with its confidence limits stated. Discuss this requirement.