Bamboo forest extent and phenological cycle mapping
Bamboo covers an estimated 35 million hectares globally yet is routinely absorbed into broadleaf forest classes in national inventories, distorting carbon accounts and timber assessments. Dense satellite time series expose bamboo's distinctive phenological signature where standard land-cover maps see only undifferentiated canopy.
Sensors
- Sentinel-2 MSI: 10 m spatial resolution across 13 spectral bands; 5-day revisit at the equator (10 days per satellite). The red-edge bands (B5, B6, B7 at 20 m) and near-infrared (B8 at 10 m) are particularly sensitive to the rapid chlorophyll surge during bamboo leaf flush, which can raise NDVI by 0.15 to 0.25 in under three weeks.
- MODIS MOD13Q1: 250 m NDVI and EVI composites at 16-day intervals, with a continuous archive back to 2000. Coarse for stand-level mapping but essential for characterising multi-year phenological cycles, including the irregular mast-seeding and mass die-off events that can span one to three years and are visible as a sustained NDVI collapse across a stand.
- Sentinel-1 SAR (C-band): 6-day repeat in IW mode at 10 m resolution. C-band backscatter distinguishes bamboo from grassland and low scrub in persistently cloudy regions where optical time series have large gaps. Bamboo culm density and hollow stem structure produce a characteristic cross-polarisation (VH) response that differs from both grass and broadleaf canopy, though the signal overlaps with some dense reed-bed communities.
- Landsat 8/9 OLI: 30 m resolution, 16-day revisit per satellite (8 days combined). The long Landsat archive from 1984 onward allows reconstruction of historical bamboo extent and detection of past die-off cycles. OLI's shortwave infrared bands (SWIR1 at 1.61 µm, SWIR2 at 2.20 µm) help separate bamboo from evergreen broadleaf forest during senescence phases when bamboo moisture content drops sharply.
Why bamboo defeats standard land-cover classifiers
Most global and national land-cover products, including GlobCover, ESA CCI Land Cover and many REDD+ reference maps, assign bamboo-dominated stands to broadleaf forest or mixed forest classes. The confusion is understandable: at a single point in time, a mature bamboo stand has a closed canopy, high leaf-area index and near-infrared reflectance that sits squarely within the broadleaf forest envelope. A single Sentinel-2 scene, however good its resolution, cannot resolve the ambiguity.
The distinguishing information is in time, not in any one spectral snapshot. Bamboo species in the genera Phyllostachys, Moso, Dendrocalamus and Guadua follow growth rhythms that no broadleaf tree matches: new culms can extend several centimetres per day during the spring shooting phase, producing a rapid and spatially coherent NDVI rise across an entire stand. Broadleaf forests show seasonal greenup too, but the rate and synchrony of bamboo flush is measurably different when you have enough observations. Moso bamboo in southern China, for instance, shows a characteristic double-peak NDVI pattern tied to its biennial shooting cycle, documented in published studies using MODIS time series. That double-peak is absent in co-located broadleaf forest.
The mast-seeding signal: a catastrophic phenology that shows up from orbit
Many bamboo species are monocarpic: they flower once, set seed en masse across an entire population, and then die. The interval between flowering events varies by species, from roughly 15 years in some Melocanna species to 60 or more years in Phyllostachys bambusoides. When a stand enters mass flowering, the canopy does not simply decline gradually. It collapses over one to three years, producing a sustained NDVI drop that is spatially coherent across potentially thousands of hectares and is not associated with fire, logging or drought stress.
This signature is detectable in MODIS MOD13Q1 time series as an anomalous multi-year departure from the site's historical NDVI baseline. It is also visible in Landsat and Sentinel-2 archives as a progressive shift from closed canopy to bare ground or regenerating seedling cover. The same archive that records the die-off can be used to estimate when regeneration began and to project when the stand will return to closed-canopy status, which matters for carbon accounting: a post-die-off bamboo stand is a net carbon source for several years before becoming a sink again.
Building the time series: what works and what does not
The practical workflow combines Sentinel-2 surface-reflectance time series (processed through the Sen2Cor atmospheric correction chain or equivalent) with MODIS composites for phenological context. The Sentinel-2 data provide spatial detail sufficient to map stands of a few hectares; MODIS provides the long historical record needed to characterise multi-year cycles. Landsat bridges the two where Sentinel-2 archive depth is insufficient.
Cloud cover is the principal constraint in the regions where bamboo is most ecologically significant: the eastern Himalayas, southwest China, the Western Ghats, and parts of tropical South America. In Assam or Arunachal Pradesh, optical revisit can effectively drop to one usable scene per month or fewer during the monsoon. Sentinel-1 SAR fills some of that gap, but its ability to distinguish bamboo from structurally similar vegetation is limited; it is most useful as a binary forest/non-forest mask rather than a bamboo-specific classifier. Honest programme design in cloud-prone regions should plan for phenological feature extraction over a full annual cycle, accepting that some years will have sparse optical coverage.
The minimum mappable stand size at Sentinel-2 resolution is roughly 0.5 to 1 hectare for a spectrally pure bamboo patch, assuming adequate cloud-free observations during the flush period. Smaller stands or mixed bamboo-tree edges are likely to be missed or misclassified.
Carbon and inventory implications of getting the class right
The carbon accounting error from misclassifying bamboo as broadleaf forest is not trivial. Published estimates of bamboo above-ground carbon density vary widely by species and management, but the values are generally lower than those for mature tropical broadleaf forest. Assigning broadleaf forest carbon factors to bamboo stands therefore inflates national carbon stock estimates. Conversely, bamboo's rapid biomass accumulation rate means that regenerating stands can sequester carbon faster than equivalent broadleaf regrowth, a fact that is lost when the class is collapsed.
Several countries with significant bamboo resources, including China, India, Ethiopia and Brazil, have begun to treat bamboo separately in their national forest inventories and Nationally Determined Contributions under the Paris Agreement. Accurate spatial extent maps, updated at least every two to three years to capture die-off and regeneration cycles, are a prerequisite for that accounting to be credible. Satellize's analytics stack, which runs on open Sentinel and Landsat constellations, can produce those maps at national scale without requiring countries to commission new satellite infrastructure.
Spectral indices and classification features that actually discriminate
Beyond NDVI, several published spectral and temporal features improve bamboo discrimination. The red-edge chlorophyll index (CIre, computed from Sentinel-2 B7 and B5) is sensitive to the high chlorophyll content during flush and drops sharply during senescence, giving a larger seasonal amplitude than NDVI alone. The SWIR-based normalised difference moisture index tracks the pronounced moisture loss during die-off. Combining these indices into a multi-temporal feature stack, then applying a random forest or support vector machine classifier trained on field-verified bamboo polygons, is the published approach used in peer-reviewed mapping studies for Moso bamboo in China and Guadua in the Andes.
The key classifier inputs are not individual band values but phenological metrics derived from the time series: the date and rate of the spring NDVI rise, the amplitude of the annual cycle, the presence of the biennial double-peak in Moso, and the long-term trend. These features are stable across years in established stands and change dramatically during die-off, making them useful for both initial mapping and change detection. Accuracy in published studies for Moso bamboo mapping in Zhejiang and Fujian provinces has reached overall accuracies above 85 percent at 10 to 30 m resolution, though performance degrades in fragmented landscapes where bamboo and broadleaf forest are finely intermixed.
Typical figures
| Primary optical resolution | 10 m (Sentinel-2 NIR), 20 m (Sentinel-2 red-edge), 30 m (Landsat OLI) |
| SAR resolution | 10 m (Sentinel-1 IW mode, ground range) |
| Optical revisit (cloud-free target) | 5 days at equator (Sentinel-2 combined); 8 days (Landsat 8+9 combined); 16-day composites (MODIS MOD13Q1) |
| SAR revisit | 6 days (Sentinel-1, single satellite) |
| Minimum mappable bamboo stand | 0.5 to 1 ha (Sentinel-2); ~6 ha practical threshold (MODIS alone) |
| Key spectral bands | NIR (842 nm), red-edge (705, 740, 783 nm), SWIR1 (1610 nm), SWIR2 (2190 nm); C-band SAR VV/VH polarisation |
| Phenological cycle detection archive depth | Sentinel-2 from 2017; Landsat from 1984; MODIS from 2000 |
| Latency (map update) | 2 to 4 weeks after end of target phenological window, subject to cloud clearance |
| Typical classification accuracy (published range) | 82 to 91% overall accuracy for Moso bamboo in published studies; lower in fragmented or cloud-prone landscapes |
| Delivery formats | GeoTIFF (classified extent), GeoPackage (stand polygons with phenological attributes), CSV (per-stand phenological metrics) |
Analytics Satellize can run
| National bamboo extent map | Multi-temporal random forest classification on Sentinel-2 phenological feature stack (NDVI amplitude, CIre seasonal range, SWIR moisture index, flush-onset date) | GeoTIFF and polygon layer with bamboo/non-bamboo class and confidence score per pixel |
| Phenological cycle characterisation per stand | Harmonic time-series decomposition (Fourier fitting) on MODIS MOD13Q1 and Sentinel-2 NDVI stacks; biennial double-peak detection for Moso-type species | Per-stand CSV with flush date, amplitude, cycle period estimate and anomaly flag |
| Mass die-off detection and area estimate | NDVI breakpoint detection (BFAST or equivalent) on MODIS and Landsat long-term archive; sustained multi-year decline flagged against site-specific historical baseline | Polygon layer of die-off areas with onset year, affected area (ha) and regeneration status |
| Bamboo vs. grassland discrimination in cloudy regions | Sentinel-1 C-band VH backscatter time series combined with available optical observations; threshold and machine-learning fusion to separate bamboo from grass and scrub | Binary raster mask (bamboo / non-bamboo) with per-pixel SAR confidence score |
| Historical extent reconstruction | Landsat OLI archive classification (1984 onward) using phenological features where multi-year stacks are available; gap-filled with MODIS for pre-2013 periods | Decadal extent layers in GeoTIFF; area change table by administrative unit |
| Carbon class correction for national inventory | Reclassification of existing broadleaf forest polygons using bamboo extent layer; area-weighted substitution of published bamboo carbon density factors for broadleaf factors | Revised carbon stock table (tC/ha by stand) with uncertainty range; GIS layer for submission to national forest inventory |
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.