Registry-grade satellite time-series evidence packages for carbon auditors
Carbon registries demand reproducible, provenance-tagged satellite evidence across crediting periods that can exceed three decades. Building that record requires radiometric consistency across mismatched sensors, rigorous chain-of-custody metadata, and archival strategies that outlast any single data provider.
Sensors
- Landsat 5/7/8/9 TM/ETM+/OLI: 30 m multispectral resolution, 16-day revisit per satellite (8-day with constellation pairing). Archive extends to 1972 for Landsat 1 and to 1984 for TM, giving the deepest publicly available surface-reflectance record. USGS Collection 2 Level-2 products are radiometrically normalised and atmospherically corrected to a common standard, making cross-sensor consistency tractable. Essential for establishing pre-project baselines stretching back decades.
- Sentinel-2 MSI: 10 m (visible/NIR) and 20 m (red-edge/SWIR) resolution, 5-day revisit at the equator with both satellites. ESA's Sen2Cor and the Copernicus Level-2A surface-reflectance product provide atmospherically corrected data consistent with Landsat through cross-calibration studies. Archive begins 2015. High spatial detail supports plot-level boundary verification and canopy-change attribution.
- Planet SuperDove: 3–4 m resolution, near-daily revisit globally. Eight spectral bands including red-edge. Useful for filling temporal gaps in Sentinel-2 and Landsat records during cloud-heavy seasons, and for resolving ambiguities at project boundaries. Requires explicit radiometric harmonisation to Sentinel-2 or Landsat reference scenes before inclusion in a normalised time series; Planet publishes surface-reflectance products but inter-sensor calibration residuals must be documented.
- Sentinel-1 SAR (C-band): 10 m resolution in Interferometric Wide swath mode, 6–12 day revisit. Cloud-penetrating, so it provides acquisition continuity during the wet-season periods when optical sensors are most frequently obscured. Backscatter time series can detect canopy disturbance independently of optical data, strengthening evidence packages for projects in persistently cloudy regions. Not a substitute for spectral reflectance but a corroborating layer.
Why a time series is not the same as a collection of images
A carbon auditor reviewing a REDD+ project does not want a folder of GeoTIFFs. They want a consistent reflectance record in which a pixel value on one date is directly comparable to the same pixel on every other date, regardless of which satellite acquired it, which atmospheric correction version was applied, or whether a sensor was replaced mid-project. That comparability is not automatic. Landsat 7's scan-line corrector failed in 2003, leaving data gaps. Landsat 5 and 8 share no overlapping operational period. Sentinel-2A and 2B have small but measurable inter-satellite calibration offsets. Planet SuperDove generations differ in band centre wavelengths. Without explicit normalisation, apparent reflectance changes in the time series may reflect sensor history rather than forest change.
The standard approach is to select a single well-characterised sensor as the radiometric reference, typically Landsat 8 or 9 OLI using USGS Collection 2 Level-2 surface reflectance, then harmonise all other sensors to it using pseudo-invariant calibration sites or regression against contemporaneous overlapping acquisitions. ESA's Sentinel-2 harmonised products (Sentinel-2 Level-2A) are already cross-calibrated to Landsat OLI through the Copernicus programme's quality assurance framework. The residual uncertainty after harmonisation is typically quoted in the literature at a few percent of reflectance, which must be propagated into any uncertainty budget submitted to a registry.
Provenance metadata: the chain of custody a registry can actually audit
Verra's VM0007 and VM0015 methodologies, Gold Standard's Land Use and Forests framework, and the Article 6.4 supervisory body's emerging technical guidance all require that monitoring data be traceable to its source. For satellite evidence, that means each pixel observation in the time series must carry a documented lineage: the satellite identifier, acquisition date and time (UTC), orbit number or scene ID, processing baseline version, atmospheric correction algorithm and parameters, and any cloud or shadow masking flags applied. This is not bureaucratic overhead. It is what allows a third-party auditor to reproduce the analysis independently, and what allows the evidence to remain verifiable when the original analyst is no longer available.
In practice, this means building evidence packages around immutable scene identifiers from public archives. USGS Landsat Collection 2 scene IDs encode the satellite, sensor, path, row, acquisition date and processing level in a standardised string. ESA's Copernicus Dataspace assigns persistent product identifiers to every Sentinel-2 granule. These identifiers, stored alongside derived indices and classification outputs, form the audit trail. Any reprocessing, such as a change in atmospheric correction version, must be logged as a new processing event rather than silently overwriting prior outputs.
The thirty-year archive problem
A forest carbon project with a 30-year crediting period will outlast the operational life of every satellite currently in orbit, several commercial data providers, and possibly the software libraries used to process the original imagery. This is not hypothetical: Landsat 4 and 5 data was nearly lost before USGS completed its archive digitisation. The EROS archive now holds Landsat data back to 1972 on tape and disk with redundancy, and the Copernicus programme has committed to long-term data preservation, but neither commitment extends unconditionally to 2055.
A registry-grade evidence package must therefore be designed for archival independence. The minimum viable approach stores the processed, normalised reflectance time series alongside its provenance metadata in open formats (GeoTIFF or NetCDF with embedded metadata, not proprietary database exports), held in at least two geographically separated repositories. The raw scene IDs must be preserved so that any future auditor can re-derive the analysis from the public archive if the processed outputs are questioned. Where commercial data such as Planet imagery is included, the licence terms must permit archival retention and third-party audit access for the full crediting period, a requirement that is not always met by standard commercial licences and must be negotiated explicitly.
Cloud cover: the honest accounting
Optical sensors cannot see through cloud. In persistently cloudy tropical regions, a 16-day Landsat revisit may yield fewer than four usable acquisitions per year. Sentinel-2's 5-day revisit improves the odds but does not eliminate the problem. The practical consequence is that annual monitoring periods, which most methodologies require, may rest on a small number of clear observations unevenly distributed across the year. This matters because phenological variation, dry-season stress browning in particular, can mimic disturbance signals if the clear observations happen to cluster in a single season.
Honest evidence packages document the temporal distribution of clear observations explicitly, flag periods where cloud cover exceeded a stated threshold (commonly 20–30% of the project area), and use Sentinel-1 SAR backscatter as a supplementary disturbance indicator during those periods. SAR cannot replicate spectral indices such as NDVI or NBR, but it can confirm whether a large-scale canopy removal event occurred during an optically obscured window. Registries increasingly accept multi-sensor fusion approaches, but the fusion logic and its uncertainty implications must be documented, not assumed.
Constructing the deliverable an auditor can sign off
A registry-grade evidence package is a structured archive, not a report. Its core components are: a normalised surface-reflectance time series in open format with embedded provenance; derived spectral index stacks (NDVI, NBR, SWIR-based disturbance indices) computed from that reflectance; a cloud and shadow mask layer for every acquisition date; a change detection output with per-pixel confidence scores and the algorithm version used; and a human-readable processing log that maps every output to its input scenes by their public archive identifiers.
The package should be versioned. If a methodology update or a reprocessing event changes any output, the prior version must be retained and the change documented. This version control discipline is standard in software engineering but rare in remote-sensing practice. Adopting it from the outset is far cheaper than reconstructing an audit trail retrospectively when a verification body raises a query five years into a crediting period.
Satellize builds evidence packages of this type on open constellations (Sentinel, Landsat) with commercial gap-fill on client licence, applying the same provenance discipline it uses in its Tonga crop-estimation programme. Auditors who want to understand the processing chain before commissioning a package can request a technical specification review as a first step.
Typical figures
| Spatial resolution (optical) | 10–30 m (Sentinel-2 / Landsat); 3–4 m with Planet SuperDove gap-fill |
| Spatial resolution (SAR) | 10 m (Sentinel-1 IW mode) |
| Revisit frequency | 5 days (Sentinel-2 constellation); 8–16 days (Landsat 8+9 pair); near-daily (Planet) |
| Archive depth | 1984 to present (Landsat TM/ETM+/OLI, USGS Collection 2); 2015 to present (Sentinel-2) |
| Spectral bands used | Blue, Green, Red, NIR, Red-edge, SWIR1, SWIR2 (optical); C-band VV/VH (SAR) |
| Radiometric standard | Surface reflectance (USGS Collection 2 Level-2; ESA Sentinel-2 Level-2A) |
| Minimum detectable disturbance | Canopy clearance of approximately 0.1 ha reliably detectable at 10 m resolution; smaller patches carry higher false-negative risk |
| Provenance metadata | Scene ID, satellite, acquisition UTC, orbit/path/row, processing baseline version, cloud mask version — per observation |
| Delivery formats | Cloud-optimised GeoTIFF (COG), NetCDF-4 with CF metadata, STAC-compliant catalogue JSON |
| Archival retention design | Open formats in two geographically separated repositories; raw scene IDs preserved for independent re-derivation |
Analytics Satellize can run
| Cross-sensor normalised reflectance time series | Pseudo-invariant feature site regression and/or empirical line normalisation to Landsat OLI Collection 2 reference; residual uncertainty propagated and documented | Cloud-optimised GeoTIFF stack per spectral band, one layer per acquisition date, with embedded provenance metadata |
| Annual spectral index composites (NDVI, NBR, SWIR disturbance index) | Medoid compositing over cloud-free observations within each annual monitoring window, following methods consistent with Landsat-based compositing literature | Annual composite GeoTIFF layers with per-pixel observation count and date-range fields |
| Cloud and shadow fraction report per monitoring period | Fmask or equivalent scene-level cloud/shadow classification applied to every acquisition; temporal gap analysis flagging periods below usability threshold | Per-period tabular report and spatial mask layers; flags periods requiring SAR supplementation |
| Canopy disturbance change map with confidence scores | Continuous Change Detection and Classification (CCDC) or LandTrendr applied to normalised time series; per-pixel change magnitude, date and confidence score | GIS polygon layer with disturbance attributes; algorithm version and parameter set recorded in processing log |
| SAR-based disturbance corroboration layer | Sentinel-1 C-band backscatter time series analysis (VV/VH ratio change detection) for cloud-obscured periods | Binary disturbance indicator raster for each optically obscured monitoring window, with acquisition date range metadata |
| Audit-ready evidence package assembly | Structured archive construction: versioned outputs, immutable scene ID index, processing log, STAC catalogue linking every output to its source scenes | Versioned archive in open formats with human-readable processing log; suitable for submission to Verra, Gold Standard or Article 6 verification bodies |
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.