Tropical secondary forest age-class mapping for baseline setting
Dense Landsat and Sentinel-2 time-series stacks can reconstruct when tropical forest was cleared and when regrowth began, pixel by pixel. The method is well-established but carries real uncertainty for older stands and cloud-persistent landscapes.
Sensors
- Landsat 4–9 (USGS/NASA): 30 m spatial resolution, 16-day revisit per satellite. The archive runs from 1982 (Landsat 4) to present, giving over 40 years of surface-reflectance observations. This depth is the foundation of any credible clearing-and-regrowth chronology in the tropics.
- Sentinel-2 MSI (ESA Copernicus): 10–20 m resolution depending on band, 5-day revisit at the equator with both satellites. Improves age-class precision for regrowth events after 2015 and helps resolve spatial detail that Landsat conflates in fragmented landscapes.
- ALOS PALSAR / PALSAR-2 (JAXA): L-band SAR at roughly 10–25 m resolution. Cloud-penetrating by design, which matters enormously in persistently overcast regions. Backscatter and coherence respond to vegetation structure rather than greenness, offering a complementary signal for early-stage woody regrowth.
- MODIS Terra/Aqua (NASA): 250–500 m resolution, daily revisit. Too coarse to map individual clearing events in fragmented landscapes, but useful for establishing long-baseline NDVI trajectories before Landsat 4 observations are available and for cross-checking spectral trends.
Why stand age is not just a number
Carbon additionality in avoided-deforestation and reforestation projects depends on demonstrating that a forest is younger, and therefore accumulating carbon faster, than the baseline would assume. Tropical secondary forests accumulate above-ground biomass rapidly in early decades: published pantropical studies (Poorter et al., 2016 in Nature; Heinrich et al., 2021 in Science) document median above-ground biomass recovery of roughly 50 percent of old-growth levels within 20 years, though the range across sites is wide. A project claiming additionality for 10-year-old regrowth is making a very different carbon argument than one claiming it for 30-year-old regrowth. Auditors and registry methodologies increasingly require pixel-level evidence of when clearing occurred and when canopy closure began.
The problem is that tropical landscapes are dynamic and the historical record is incomplete. Clearing events, burn scars, agricultural use, and spontaneous regrowth leave overlapping spectral footprints. Untangling them requires reading the full archive, not a single image pair.
What a spectral trajectory gives away
LandTrendr (Landsat-based Detection of Trends in Disturbance and Recovery), developed at Oregon State University and published by Kennedy et al. in Remote Sensing of Environment, fits segmented linear models to annual spectral index time-series at each pixel. The algorithm identifies abrupt drops in vegetation indices (clearing, burning) and subsequent recovery trajectories, assigning a year of disturbance and a rate of recovery. Applied to a full Landsat archive stack, it can date clearing events to within one to two years for most of the record from the mid-1980s onward.
The NBR (Normalised Burn Ratio) and NDVI are the most commonly used indices. NBR is particularly sensitive to canopy structure loss and recovery. For Sentinel-2 data from 2015 onward, the same trajectory logic applies at finer spatial resolution, refining age-class boundaries in fragmented landscapes where 30 m Landsat pixels mix forest edge with clearing.
PALSAR L-band backscatter adds a structurally independent signal. Young regrowth (under roughly five years) tends to show low HV backscatter; woody biomass accumulation increases it measurably. In cloud-persistent regions where optical time-series have large gaps, PALSAR coherence and backscatter can anchor the trajectory where Landsat observations are sparse.
Where the uncertainty actually lives
Honest age-class mapping requires stating where the method degrades. Three zones of genuine uncertainty exist.
First, older secondary stands. Spectral indices in secondary forest typically converge toward old-growth values somewhere between 15 and 25 years post-clearing, depending on site. After that convergence, optical data alone cannot distinguish a 25-year-old stand from a 40-year-old one. The trajectory signal is gone. Projects claiming additionality for stands older than roughly two decades should treat the optical age estimate as a lower bound and seek corroborating evidence from PALSAR structural data, field plots, or historical aerial photography where it exists.
Second, cloud-persistent regions. The western Amazon, much of the Congo Basin, and parts of insular Southeast Asia can have fewer than four to six usable Landsat observations per year. Sparse time-series increases the probability that a clearing event falls entirely between cloud-free acquisitions, shifting the estimated disturbance date by one or more years. Uncertainty bands should widen accordingly, and PALSAR gap-filling becomes more important.
Third, repeated disturbance. Shifting cultivation cycles leave multiple overlapping disturbance signals. LandTrendr can detect multiple segments, but interpreting them as distinct agricultural cycles versus selective logging versus natural disturbance requires ancillary data (land-use records, local knowledge) that remote sensing alone cannot supply.
Building the age-class map: from archive to deliverable
The practical workflow begins with surface-reflectance preprocessing: cloud masking, atmospheric correction, and pixel-quality filtering using the CFMask algorithm embedded in Landsat Collection 2 products and the Sen2Cor processor for Sentinel-2. Annual composites are then built from the best available cloud-free observations within each calendar year, typically favouring dry-season windows where they exist.
LandTrendr or equivalent change-detection algorithms (CCDC, Continuous Change Detection and Classification, developed by Zhu and Woodcock, is a published alternative that operates on all available observations rather than annual composites) are then run across the full stack. The output is a per-pixel disturbance year, a recovery rate parameter, and a confidence score derived from the number of clean observations supporting each segment.
Age-class maps are then binned: commonly 0–5 years, 6–10, 11–20, 21–30, and greater than 30 years post-disturbance, though project methodologies vary. Each class carries a different biomass accumulation trajectory drawn from published allometric and chronosequence studies for the relevant biome. The map, the underlying trajectory stack, and the per-pixel uncertainty layer together constitute the evidentiary package a carbon registry auditor can interrogate.
What this means for a baseline audit
A credible baseline requires demonstrating that the project area was not already recovering forest at the time the project was registered. Age-class mapping is the primary instrument for that demonstration. If the trajectory analysis shows that canopy recovery began five years before the project start date, the additionality claim for those pixels is weakened or eliminated.
The same analysis works in reverse for leakage buffers: if clearing events appear in the buffer zone after project registration, their spectral trajectories can be dated and spatially attributed. Satellize applies this trajectory-analysis workflow on open archive data (Landsat Collection 2, Sentinel-2 Level-2A) for carbon project developers and auditors who need a defensible, reproducible record rather than a one-time snapshot.
One practical note on delivery format: the evidentiary value of a trajectory analysis is only as good as its reproducibility. Delivering a GIS layer without the underlying annual composites and algorithm parameters is not sufficient for registry-grade audit. The full processing chain, including cloud-mask logs and algorithm version, should accompany any age-class map submitted as MRV evidence.
Typical figures
| Spatial resolution (optical) | 30 m (Landsat 4–9); 10–20 m (Sentinel-2 MSI) |
| Spatial resolution (SAR supplement) | 10–25 m (ALOS PALSAR-2 Fine Beam) |
| Revisit frequency | 16 days per Landsat satellite; ~5 days (Sentinel-2 A+B combined at equator) |
| Archive depth (optical) | Landsat: 1982–present (40+ years); Sentinel-2: 2015–present |
| Usable observations per year (cloud-persistent tropics) | As few as 4–6 cloud-free Landsat scenes annually in western Amazon or Congo Basin; Sentinel-2 improves density post-2015 but does not eliminate the problem |
| Age-class dating precision | ±1–2 years for disturbance events with adequate cloud-free observations; widens to ±3–5 years in data-sparse years or cloud-persistent regions |
| Minimum clearing size reliably detected | ~0.1 ha at Sentinel-2 resolution; ~0.5 ha at Landsat 30 m (sub-pixel mixing limits detection below these thresholds) |
| Spectral indices used | NBR, NDVI, SWIR-based indices (e.g. NBR2); L-band HV backscatter for SAR |
| Deliverable formats | GeoTIFF age-class raster, per-pixel uncertainty layer, annual composite stack, change-segment vector polygons, processing metadata log |
| Applicable change-detection algorithms | LandTrendr (Kennedy et al.); CCDC (Zhu & Woodcock); BFAST (Verbesselt et al.) |
Analytics Satellize can run
| Pixel-level disturbance chronology map | LandTrendr spectral segmentation on annual Landsat NBR composites, with Sentinel-2 refinement post-2015 | GeoTIFF raster: year of most recent clearing event per pixel, with confidence score layer |
| Secondary forest age-class map | Time-since-disturbance calculation from disturbance chronology, binned into project-specified age classes | GeoTIFF raster and vector polygon layer with area statistics per age class, suitable for registry submission |
| Repeated-disturbance history layer | Multi-segment LandTrendr or CCDC analysis identifying pixels with more than one clearing event in the archive | Vector layer flagging pixels with complex histories; narrative annex for auditor review |
| Cloud-gap uncertainty assessment | Per-pixel count of usable observations per year; identification of years where fewer than three cloud-free scenes exist | Uncertainty raster and tabular summary by project sub-zone, with recommended PALSAR supplementation zones |
| PALSAR-supplemented recovery index | L-band HV backscatter trajectory analysis (JAXA PALSAR/PALSAR-2 mosaics) fused with optical age-class estimates in cloud-persistent zones | Fused age-class GeoTIFF with source-flag layer indicating optical-only versus SAR-supplemented pixels |
| Pre-project baseline age-class snapshot | Age-class map frozen at project registration date, derived from archive trajectory to that date | Timestamped GeoTIFF and summary statistics report formatted for Verra VCS or Gold Standard baseline documentation |
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.