Dryland forest carbon stock adjustment for water-stress mortality

Drought-driven canopy dieback in semi-arid woodland carbon projects silently erodes biomass stocks while baseline estimates stay frozen. Combining Sentinel-2 NDVI anomaly time-series with Landsat 8/9 thermal data lets project auditors catch stress-driven reversals before they become registry liabilities.

Sensors

Sentinel-2 MSI: 10 m resolution in visible and near-infrared bands (B4, B8) used to compute NDVI; 20 m in red-edge bands (B5, B6, B7) useful for early stress detection before full dieback. Five-day revisit at the equator under cloud-free conditions. Free archive from 2015.
Landsat 8/9 TIRS: Thermal infrared bands 10 and 11 at 100 m native resolution (resampled to 30 m in products) retrieve land-surface temperature. Elevated LST over stressed canopy, relative to surrounding healthy woodland, is a reliable stress proxy. Sixteen-day revisit per satellite; combined Landsat 8 and 9 gives eight-day effective revisit.
MODIS Terra/Aqua MOD13 vegetation products: 250 m NDVI and EVI composites at 16-day intervals provide a long historical baseline stretching back to 2000, essential for computing multi-year NDVI anomalies in dryland systems where inter-annual rainfall variance is high. Coarse resolution makes MODIS unsuitable for plot-level accounting but ideal for contextual drought characterisation.
ECMWF ERA5 soil-moisture reanalysis: Gridded soil-moisture and evapotranspiration deficit fields at roughly 31 km horizontal resolution, available from 1940 to near-real-time. Used to attribute NDVI anomalies to water stress rather than phenological variation, and to flag multi-season drought accumulation that precedes visible dieback.

Why dryland carbon projects carry a hidden reversal risk

Semi-arid and sub-humid woodlands, think miombo, mulga, savanna woodland and the drier end of Mediterranean shrubland, are increasingly popular carbon project hosts. Their biomass density is modest but the areas are large, additionality arguments are often credible, and land costs are low. The problem is that these ecosystems sit close to a physiological cliff. A single severe drought, or two moderate ones in succession, can kill a meaningful fraction of the standing stock. Trees that die do not immediately fall or decompose; they stand for years, invisible to a ground crew that visits once a season.

Registry methodologies, including Verra's VM0007 and VM0009 frameworks, require projects to account for reversals when carbon stocks decline below baseline. The obligation is clear. The detection mechanism often is not. Without systematic satellite monitoring, dieback can accumulate across thousands of hectares before a field audit catches it, by which time the project may already have issued credits against stock that no longer exists.

What a dying canopy looks like from orbit

A healthy tree canopy reflects strongly in the near-infrared (around 700 to 900 nm) and absorbs in the red (around 665 nm). NDVI, the normalised ratio of those two signals, sits above 0.4 for vigorous dryland woodland and drops sharply as leaf water content falls and chlorophyll degrades. Sentinel-2's red-edge bands at 705 nm and 740 nm are particularly sensitive to early chlorophyll stress, often flagging physiological decline four to eight weeks before broadband NDVI collapses visibly. That lead time matters: it is the difference between a project manager who can investigate and one who is already in deficit.

Thermal data adds a second, independent line of evidence. A stressed canopy transpires less, so its surface temperature rises relative to healthy neighbours. Landsat TIRS land-surface temperature retrievals over dying woodland patches typically run two to five degrees Celsius above background, a signal detectable even at 100 m native resolution when the affected patch exceeds roughly two to three hectares. Below that spatial scale, the thermal signal is diluted by surrounding healthy canopy and the retrieval becomes unreliable. That is an honest limit worth stating clearly.

Building an anomaly time-series that can survive a registry audit

A single low-NDVI image proves nothing. Dryland vegetation is phenologically noisy: leaf flush and senescence track rainfall pulses, not calendar months, and a dry year can produce NDVI values that look alarming but represent normal dormancy. The analytical approach that holds up to scrutiny compares each image against a percentile envelope built from the same calendar period across multiple prior years, typically the full MODIS record from 2000 and the Sentinel-2 archive from 2015. A pixel that sits below the 10th percentile of its historical distribution for three or more consecutive compositing periods is flagged as anomalous. Persistent anomaly, sustained across two or more growing seasons, is the threshold for classifying mortality rather than temporary stress.

ERA5 soil-moisture deficit fields provide the meteorological context that makes the anomaly interpretable. If a cluster of low-NDVI pixels coincides with a documented multi-month soil-moisture deficit in the ERA5 record, the drought attribution is defensible. If the anomaly appears without a corresponding moisture signal, other causes, pest outbreak, fire scorch, illegal clearing, need investigation. The satellite stack cannot resolve that ambiguity on its own. That is not a flaw in the method; it is a boundary condition that honest MRV documentation should state explicitly.

The 10-metre floor and what falls below it

Sentinel-2's 10 m resolution sounds fine-grained, but in open dryland woodland with canopy cover of 20 to 40 percent, individual tree crowns are often two to six metres across. A pixel at 10 m integrates canopy and bare soil in proportions that shift with sun angle and canopy architecture. NDVI computed from mixed pixels underestimates the severity of mortality in sparse stands and can miss the death of individual large trees entirely unless they form contiguous patches of at least two to three pixels.

This resolution floor has a practical consequence for stock adjustment. Mortality estimates derived from 10 m optical data should be treated as lower bounds in sparse woodland. Projects in very open savanna, where canopy cover drops below 20 percent, may need to supplement with commercial very-high-resolution imagery, or accept wider uncertainty bounds in the carbon accounting. The thermal channel from Landsat TIRS, at 100 m native resolution, is even coarser and is best used as a screening tool to prioritise field verification rather than as a primary mortality mapping layer.

Translating dieback maps into carbon stock adjustments

Mortality mapping produces a spatial layer: pixels classified as stressed, severely stressed, or dead, with confidence scores derived from the persistence and depth of the NDVI anomaly. Converting that layer into a carbon stock adjustment requires a biomass model specific to the project's woodland type, typically an allometric equation relating canopy area or crown diameter to above-ground dry mass. Published allometrics for miombo, acacia and eucalypt woodland are available in the peer-reviewed literature and in the FAO Global Forest Resources Assessment supporting data. Satellize does not fabricate species-specific allometrics; the analytics pipeline ingests the project's own validated equations.

The output is a revised above-ground carbon stock estimate for the affected area, with uncertainty propagated from both the remote-sensing classification and the allometric model. That figure feeds directly into the buffer pool calculation that registries require when reversals occur. Producing it before the reporting deadline, rather than after a field audit discovers the problem, is the operational point of the exercise. Satellize runs this pipeline on open constellations for clients who need registry-grade documentation; the Tonga crop-estimation programme uses a structurally similar anomaly-detection approach applied to agricultural rather than woodland canopy.

Scheduling the monitoring cycle around reporting windows

Most carbon registries require annual or biennial verification reports. A monitoring cycle that produces one cloud-free composite per growing season is the minimum viable cadence for dryland projects. In practice, cloud contamination in the wet season, when vegetation is most active, can reduce usable Sentinel-2 acquisitions to three or four per year across parts of southern Africa and northern Australia. Scheduling the primary NDVI anomaly assessment in the early dry season, when cloud cover is low and stress signals are already visible in the canopy, gives the best chance of a clean, audit-ready image stack.

A practical workflow runs MODIS composites monthly for near-real-time screening, triggers a Sentinel-2 and Landsat analysis when MODIS flags a persistent anomaly, and then schedules field verification for any patch that exceeds a project-defined mortality threshold. The field team confirms whether the cause is drought, pest or something else. That confirmation is recorded in the monitoring report alongside the satellite evidence. The result is a defensible, multi-source record rather than a single contested data point.

Typical figures

Primary optical resolution	10 m (Sentinel-2 B4/B8 NDVI); 20 m (Sentinel-2 red-edge B5/B6/B7)
Thermal resolution	100 m native, 30 m resampled (Landsat 8/9 TIRS)
Revisit cadence	5 days (Sentinel-2 at equator); 8 days effective (Landsat 8 + 9 combined); 16-day composites (MODIS MOD13)
Contextual soil-moisture grid	~31 km horizontal resolution, ERA5 reanalysis; latency approximately 5 days to near-real-time
Minimum detectable mortality patch	Approximately 0.03 ha (3 × 10 m pixels) for optical classification; approximately 2–3 ha for thermal confirmation
NDVI anomaly archive depth	MODIS from 2000; Sentinel-2 from 2015; Landsat thermal from 2013 (Landsat 8)
Spectral bands used	Red (665 nm), NIR (842 nm), red-edge (705, 740, 783 nm) for optical; 10.6–12.5 µm for thermal
Latency from acquisition to anomaly flag	Typically 3–7 days for cloud-screened composite processing; longer during persistent cloud cover
Deliverable formats	GeoTIFF mortality classification rasters, GeoJSON anomaly polygons, tabular carbon stock adjustment estimates, PDF monitoring report annexes

Analytics Satellize can run

NDVI anomaly percentile map	Per-pixel historical percentile envelope from MODIS and Sentinel-2 archive; anomaly defined as sustained departure below 10th percentile	Seasonal GeoTIFF raster with per-pixel anomaly severity score, clipped to project boundary
Canopy mortality classification layer	Multi-temporal thresholding of NDVI anomaly persistence across two or more growing seasons, validated against Landsat LST elevation	GeoJSON polygon layer with mortality confidence class (stressed, severely stressed, dead) and area statistics
Land-surface temperature anomaly map	Landsat TIRS split-window LST retrieval compared against multi-year baseline for same calendar period	GeoTIFF LST anomaly layer (degrees Celsius above baseline) for field prioritisation
Drought attribution report	Spatial correlation of NDVI anomaly clusters with ERA5 soil-moisture deficit fields; flags patches where anomaly lacks meteorological explanation	PDF report section with ERA5 time-series plots and anomaly-attribution confidence statement for registry submission
Carbon stock reversal estimate	Mortality area from classification layer applied to project-supplied allometric biomass equations; uncertainty propagated from classification accuracy and allometric variance	Tabular carbon stock adjustment (tCO₂e) with confidence interval, formatted for Verra or Gold Standard buffer pool reporting
Monitoring cycle alert feed	MODIS MOD13 monthly composite screening with threshold trigger for Sentinel-2 and Landsat tasking	Automated alert (email or API) when project-area NDVI drops below defined threshold, with link to updated raster

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.