Sensors
- Landsat 5/7/8/9 TM/ETM+/OLI: 30-metre multispectral resolution, 16-day revisit per satellite, continuous archive from 1984. The backbone of any historical baseline reconstruction; NDVI and NBR time-series from this archive are the primary evidence for testing registered deforestation trajectories.
- Sentinel-2 MSI: 10-metre resolution in visible and near-infrared bands, 5-day revisit at the equator with both satellites. Used to cross-validate Landsat-derived forest-loss polygons from 2015 onward and to sharpen boundary delineation of the reference region.
- MODIS Terra/Aqua: 250-metre resolution in bands 1 and 2, daily revisit. Useful for detecting large-scale deforestation pulses back to 2000 and for checking whether the project's counterfactual rate is consistent with regional MODIS-based forest-loss products.
What the 2023 investigations actually found
A series of investigations published in early 2023, drawing on work by researchers at the University of Cambridge and others, examined dozens of large REDD+ projects certified under the Verra Verified Carbon Standard. The central finding was not that deforestation had occurred inside project boundaries, but that the counterfactual deforestation rates used to generate credits were substantially higher than what independent reconstruction of historical data could support. One analysis suggested that fewer than ten per cent of credits in the examined projects represented real emissions reductions. The mechanism was not necessarily fraud in the simple sense; it was the quieter problem of reference-region selection bias.
A project developer choosing a reference region to estimate what deforestation would have occurred without the project has considerable discretion. Draw the boundary to include areas under active agricultural pressure and the counterfactual rate rises. Exclude areas already protected by law or geography and it rises further. Because Verra's methodology permitted substantial developer discretion in this selection, and because the resulting baseline was registered once rather than continuously audited, the incentive to optimise the reference region in the developer's favour was structural. Satellite archives can reconstruct exactly what was happening in any candidate reference region in any year, making the selection choice auditable after the fact.
Forty years of evidence that cannot be amended
The Landsat programme has produced a continuous, publicly archived record of Earth's surface at 30-metre resolution since Landsat 5's launch in 1984. USGS provides free access to the full Collection 2 archive, atmospherically corrected and geometrically consistent across sensors. For a REDD+ project registered in, say, 2010 with a ten-year historical reference period of 2000 to 2009, the auditor has access to every Landsat scene over the reference region for that entire period, plus the preceding sixteen years. The developer had access to the same data. The question is whether the registered baseline is consistent with what the archive actually shows.
The standard reconstruction approach applies a spectral index time-series, typically the Normalised Difference Vegetation Index or the Normalised Burn Ratio, to every available cloud-free Landsat observation over the reference region. Methods such as LandTrendr, developed by Kennedy and colleagues at Oregon State University and published in peer-reviewed literature, fit piecewise linear segments to these time-series to identify the timing, magnitude and duration of forest-loss events. The output is a dated map of forest disturbance that can be compared directly against the trajectory the developer registered. Discrepancies of more than a few percentage points per year, sustained across the full reference period, are a statistical signal worth examining.
The geometry of selection bias
Reference-region drawing is where the most consequential choices are made, and it is also where satellite analysis is most diagnostic. A reference region that was genuinely representative of the counterfactual pressure on the project area should share similar biophysical characteristics, similar proximity to roads and agricultural markets, similar land-tenure history, and similar deforestation rates in the years immediately before the project's start date. If the reference region's pre-project deforestation rate was already anomalously high relative to the surrounding landscape, that is evidence of selection rather than representation.
The statistical test is straightforward in principle. Using the Hansen Global Forest Change product (University of Maryland, published annually and available through GLAD) as an independent benchmark, an auditor can compare the annual forest-loss rate in the registered reference region against a set of candidate alternative reference regions drawn by a consistent spatial rule, for instance all forest within a fixed buffer of the project boundary sharing the same soil class and road-distance band. If the registered reference region sits in the upper tail of the distribution of plausible alternatives, the probability that this occurred by chance is calculable. A reference region in the 95th percentile of deforestation rates among fifty plausible alternatives is not proof of manipulation, but it is the kind of finding that warrants a formal explanation from the developer.
Cloud cover, sensor gaps and honest limits
Tropical forest baselines are the hardest case for optical reconstruction. Cloud cover in the humid tropics can render sixty to eighty per cent of Landsat scenes unusable in the wet season, and Landsat 7's scan-line corrector failure in 2003 introduced data gaps of roughly twenty-two per cent of each scene until Landsat 8 launched in 2013. These are real constraints. A reconstruction over a heavily clouded reference region in the western Amazon between 2003 and 2012 will carry wider uncertainty than one over a drier forest in southern Brazil.
The honest approach is to report uncertainty explicitly: the number of cloud-free observations per pixel per year, the fraction of the reference region with fewer than three usable observations in any given year, and the confidence interval on the estimated annual deforestation rate. Where optical data is genuinely sparse, MODIS time-series at 250 metres can provide a coarser but cloud-penetrating cross-check, and Sentinel-1 SAR data (C-band, available from 2014) can detect large clearings through cloud, though it is less reliable for detecting selective disturbance. An audit report that presents only a point estimate for the historical deforestation rate, without an uncertainty range, should itself be treated as a red flag.
From archive reconstruction to a defensible audit report
A complete baseline audit produces three outputs. First, an independent forest-cover time-series for the reference region, derived from the Landsat archive, with per-year deforestation rates and confidence intervals. Second, a spatial comparison of the registered reference region against a systematic set of alternative regions, with the registered region's deforestation rate placed in the distribution of alternatives. Third, a consistency check between the registered baseline trajectory and independent products: the Hansen Global Forest Change annual loss layers, the PRODES deforestation monitoring system for Brazilian Amazon projects, and MODIS-derived vegetation-index trends.
Where these three outputs converge on a baseline that is plausible and internally consistent, the audit provides positive assurance. Where they diverge, the report quantifies the gap and, where possible, identifies the spatial or temporal source of the discrepancy. This is not the same as determining intent; it is determining whether the registered baseline is defensible given the public archive. Satellize runs this reconstruction pipeline on open constellations, applying the same LandTrendr-class time-series methods used in the peer-reviewed literature, and can deliver results as GIS layers with full provenance metadata suitable for submission to a validation and verification body. The Tonga crop-estimation programme demonstrates the same principle at work in a different domain: the archive does not forget what the land looked like.
Typical figures
| Primary spatial resolution | 30 m (Landsat TM/ETM+/OLI); 10 m (Sentinel-2 MSI for post-2015 cross-validation) |
| Archive depth | Landsat 5 TM from 1984; MODIS from 2000; Sentinel-2 from 2015 |
| Revisit (Landsat single satellite) | 16 days; combined Landsat 8 + 9 reduces this to approximately 8 days from 2021 |
| Minimum detectable clearing | Approximately 0.09 ha at 30 m resolution; reliable detection of sustained loss typically requires clearings above 0.5 ha given cloud and phenology noise |
| Spectral bands used | Near-infrared, shortwave infrared, red (NDVI, NBR, NBR2); Landsat OLI bands 4, 5, 6, 7 |
| Cloud-cover limitation | Humid tropics: 60–80% of scenes unusable in wet season; uncertainty widens materially below 3 usable observations per pixel per year |
| Landsat 7 data gap | Scan-line corrector failure from May 2003; approximately 22% of each scene affected until Landsat 8 launch in April 2013 |
| Independent benchmark products | Hansen Global Forest Change (UMD/GLAD, annual, 30 m); MODIS Vegetation Continuous Fields (250 m, annual from 2000) |
| Delivery format | GeoTIFF time-series layers, vector polygon change maps, PDF audit report with provenance metadata |
Analytics Satellize can run
| Historical deforestation rate reconstruction | LandTrendr-class piecewise spectral time-series segmentation on Landsat Collection 2 archive | Annual forest-loss raster stack with per-pixel confidence intervals, 1984 to present, delivered as GeoTIFF |
| Reference-region selection-bias test | Spatial sampling of alternative reference regions by biophysical and proximity covariates; percentile ranking of registered region's deforestation rate against the sample distribution | Statistical summary table and map showing registered region's position in the distribution of plausible alternatives |
| Baseline consistency check against independent products | Cross-tabulation of project-registered annual loss rates against Hansen Global Forest Change and MODIS VCF annual products | Divergence report identifying years and sub-regions where registered baseline departs from independent benchmarks by more than one standard deviation |
| Cloud-cover and data-quality assessment | Per-pixel observation density mapping from Landsat scene metadata; flagging of years with fewer than three usable observations | Data-quality GIS layer quantifying uncertainty across the reference region, suitable for inclusion in a VVB submission |
| Candidate reference-region delineation | Spatial clustering of forest pixels by road distance, slope, soil class and pre-project deforestation rate using publicly available ancillary datasets | Set of alternative reference-region polygons with summary statistics, provided as a Shapefile or GeoPackage |
| Full audit evidence package | Synthesis of time-series reconstruction, selection-bias test and consistency check into a structured report following UNFCCC MRV documentation conventions | PDF report with embedded GIS layers and full data-provenance log, formatted for submission to a carbon registry or validation body |
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.