Shifting cultivation fallow-cycle mapping for avoided-deforestation baselines
Rotational fallow misclassified as primary forest inflates REDD+ additionality claims. Multi-year SAR coherence and Landsat spectral trajectories can separate cyclical clearing from permanent deforestation, but only if the archive is long enough.
Sensors
- Sentinel-1 SAR (C-band): 10 m spatial resolution, 6-day repeat at the equator with both satellites active (12-day with one). C-band coherence decays rapidly over regrowing vegetation, making it possible to detect surface change within a single 12-day interferometric pair. Archive from 2014.
- Landsat Collection 2 ARD (Landsat 8 and 9 OLI): 30 m multispectral, 16-day revisit per satellite (8-day combined). Surface reflectance Analysis Ready Data includes NDVI, NBR and tasselled-cap transformations. Archive to 1984 supports multi-decade trajectory analysis, which is essential for documenting fallow cycles longer than five years.
- ALOS-2 PALSAR-2 (L-band): L-band (23 cm wavelength) penetrates forest canopy and responds to woody stem volume rather than leaf area. Dual-polarisation HH/HV mosaics at 25 m resolution, updated annually by JAXA. Particularly useful for distinguishing low-biomass fallow regrowth from closed-canopy secondary forest where C-band saturates.
- Planet Basemaps (3 m optical): Monthly cloud-free composites at 3 m. Not a free public archive, so use is licence-dependent. Adds spatial detail for field-boundary delineation and resolves small clearings below Landsat's 30 m pixel, though spectral depth is shallower than Landsat's eight-band OLI.
Why a single-year forest map is the wrong foundation for a baseline
Shifting cultivation is not deforestation. A community that clears two hectares, farms for one or two seasons, then allows ten years of regrowth before returning is practising a land-use system that can persist for centuries without net forest loss. The problem is that REDD+ baseline maps, often constructed from a single-year or two-year optical composite, photograph that cleared patch during its farming phase and record it as a disturbance. If the same pixel is forested again by the time the project starts, the baseline map calls it forest and the clearing history disappears. The project then claims credit for protecting land that was never at risk of permanent conversion.
The inflation is not trivial. Published analyses of tropical baseline maps, including work associated with the Global Forest Watch platform, have found that fallow cycles covering five to fifteen years are systematically misread in maps built from fewer than three years of imagery. The practical consequence is overstated additionality: the project appears to protect more carbon than it actually does, and buyers of the resulting credits are paying for a fiction.
What a long time series actually reveals
The diagnostic signature of shifting cultivation is periodicity. A pixel that clears, regrows, clears again on a cycle of four to twelve years produces a distinctive oscillation in NDVI and in SAR backscatter. Landsat's archive to 1984 is long enough to capture two or three full cycles for most tropical farming systems. JAXA's annual PALSAR mosaics, available from 2007, add L-band confirmation: woody regrowth increases HV backscatter measurably within three to five years, whereas a permanently cleared pasture stays flat.
The analytical workflow starts with dense time-series fitting. The LandTrendr algorithm, developed at Oregon State University and now integrated into the Google Earth Engine implementation used by Global Forest Watch, segments each pixel's spectral history into linear segments of growth and disturbance. A fallow pixel shows repeated short disturbance segments separated by recovery segments of similar slope. A permanently deforested pixel shows one disturbance segment followed by a flat or slowly rising trajectory that never returns to pre-disturbance spectral values. Sentinel-1 coherence adds a second independent signal: coherence between repeat passes drops sharply over bare soil and low vegetation, then recovers as canopy closes. Combining coherence time series with Landsat trajectories reduces false-positive rates substantially compared with optical data alone, particularly under persistent cloud cover.
The minimum detectable cycle and the confusion with selective logging
Honesty about limits matters here, because the limits are real. Landsat's 16-day revisit, combined with cloud cover in humid tropical regions, means that in practice a usable cloud-free observation may arrive only four to eight times per year in the wettest zones. A fallow cycle shorter than about three years can fall entirely within the noise of cloud gaps and phenological variation. The method works well for cycles of five years or more. For shorter cycles, very high revisit optical data (Planet or Sentinel-2 at 10 m, 5-day revisit) is needed, and even then the spectral contrast between one-year fallow and low-intensity grazing can be ambiguous.
Selective logging creates a structurally similar problem. A logged pixel shows a partial canopy disturbance followed by partial recovery, which can mimic a short fallow cycle in Landsat NBR trajectories. L-band SAR helps here: selective logging typically leaves large woody debris that sustains HV backscatter at intermediate levels, whereas fallow clearing removes most woody material and drops backscatter to near-bare-soil values before recovery begins. The distinction is probabilistic, not categorical. Where logging roads are present, the sibling page on logging road network extraction provides a complementary constraint, but that analysis is outside scope here.
Building a defensible baseline stratum
The practical output of this analysis is a fallow-cycle probability map: for each pixel in the project area and its reference region, an estimate of the probability that observed disturbances are cyclical rather than permanent. Pixels above a defined threshold are reclassified from 'forest at risk' to 'fallow land under shifting cultivation' and removed from the additionality calculation. The threshold is a policy choice, but the map makes the choice explicit and auditable rather than hidden inside a composite image.
Stratification by fallow probability also improves the reference region. A common baseline error is drawing the reference region from a landscape with a different ratio of shifting cultivators to permanent farmers. If the project area contains 40 percent fallow land and the reference region contains 5 percent, the reference deforestation rate is structurally higher and the additionality claim is inflated before a single tree is counted. Correcting the stratum boundaries using the fallow-cycle map is a straightforward step that significantly tightens the baseline.
Satellize applies this multi-sensor trajectory approach on open constellations, with optional commercial tasking for sub-annual gap-filling in high-cloud zones. The Tonga crop-estimation programme demonstrated that dense time-series methods transfer cleanly to small-island and fragmented-landscape contexts, where pixel-level trajectory noise is otherwise a persistent problem.
Archive depth and what registries currently accept
Verra's VM0007 and VM0015 methodologies, and the emerging ART TREES standard, all require baseline reference periods of at least ten years. Landsat's 1984 archive comfortably satisfies that requirement. The practical question is whether the analysis is submitted as a static map or as a living layer that updates as new imagery arrives. Static maps are easier to audit but become stale; dynamic layers require agreed update protocols and version control. Either way, the underlying pixel-level time series should be archived and made available to third-party auditors. Satellite data stored in analysis-ready form, with provenance metadata intact, is considerably harder to manipulate after the fact than field-survey records, which is part of the argument for making it the primary evidence layer rather than a supplementary check.
Typical figures
| Optical spatial resolution | 30 m (Landsat OLI); 10 m (Sentinel-2); 3 m (Planet Basemaps, licence required) |
| SAR spatial resolution | 10 m (Sentinel-1 IW mode); 25 m (ALOS-2 PALSAR-2 annual mosaic) |
| Optical revisit | 8 days combined Landsat 8+9; 5 days Sentinel-2A+B; daily Planet (cloud-limited in tropics) |
| SAR revisit | 6 days Sentinel-1 (dual satellite, equatorial); 12 days single satellite; ALOS-2 annual mosaic |
| Archive depth | Landsat to 1984 (40+ years); Sentinel-1 from 2014; PALSAR mosaics from 2007 |
| Minimum detectable fallow cycle | Approximately 5 years with Landsat alone in high-cloud zones; 3 years with Sentinel-2 dense stack |
| Spectral bands used | NIR, SWIR-1, SWIR-2, Red (Landsat/Sentinel-2 NDVI, NBR, tasselled-cap); C-band VV/VH (Sentinel-1); L-band HH/HV (PALSAR-2) |
| Coverage | Global tropics; Sentinel-1 coverage varies by acquisition mode and region |
| Latency (baseline product) | Baseline maps are retrospective; update cycle typically annual or per verification period |
| Delivery formats | GeoTIFF fallow-probability raster, vector stratum boundaries (GeoPackage/Shapefile), pixel-level time-series CSV for audit |
Analytics Satellize can run
| Fallow-cycle probability map | LandTrendr spectral segmentation on Landsat Collection 2 ARD dense time series, flagging pixels with two or more disturbance-recovery cycles | GeoTIFF raster with per-pixel fallow probability score and cycle-count attribute, covering project area and reference region |
| SAR coherence change stack | Sentinel-1 interferometric coherence computed over sequential 12-day pairs; temporal coherence profile fitted to distinguish bare-soil fallow from closed-canopy forest | Annual coherence summary rasters and time-series plots per stratum, formatted for inclusion in MRV technical annexes |
| L-band biomass trajectory layer | JAXA ALOS-2 PALSAR-2 annual HV backscatter time series; pixel-level slope analysis to separate recovering fallow from static low-biomass land cover | Multi-year HV backscatter stack with change-magnitude layer, GeoTIFF format |
| Revised baseline stratum boundaries | Overlay of fallow-probability map with existing project forest stratum; reclassification of pixels above agreed threshold from 'forest at risk' to 'shifting cultivation fallow' | Vector GeoPackage of corrected stratum polygons with area statistics and reclassification log for registry submission |
| Confusion matrix for fallow vs. selective logging | Combined Landsat NBR trajectory shape analysis and PALSAR-2 HV backscatter level to assign posterior probability of fallow vs. logging disturbance origin | Tabular confusion assessment with per-class accuracy estimates and written uncertainty statement suitable for third-party audit |
| Pixel-level provenance archive | Full input imagery stack with acquisition dates, cloud-mask metadata and algorithm version recorded per pixel; stored in analysis-ready format | Compressed archive (cloud-optimised GeoTIFF stack) with accompanying metadata JSON, retained for the duration of the crediting period |
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.