Multispectral mangrove extent and canopy density change detection
Mangrove forests are mapped and monitored using red-edge, NIR and SWIR spectral bands, SAR backscatter, and tidal-phase-controlled image selection. Accurate extent and canopy density change detection requires understanding where these methods break down.
Sensors
- Sentinel-2 MSI: 10 m resolution in visible and NIR bands, 20 m in red-edge (bands 5, 6, 7) and SWIR (bands 11, 12). Five-day revisit at the equator with two satellites. The red-edge bands are the primary discriminator between mangrove and adjacent terrestrial forest. Free and open archive from 2015.
- Landsat 8/9 OLI: 30 m multispectral resolution including NIR and SWIR bands, 16-day single-satellite revisit. No dedicated red-edge band, which limits canopy-type discrimination, but the archive extends to 1984 (Landsat 5 TM), making it the only source for multi-decade baseline comparisons.
- Sentinel-1 SAR (C-band): Cloud-penetrating C-band backscatter at 10 m resolution in IW mode. Sensitive to canopy volume and surface roughness. Useful for detecting structural change under persistent cloud cover, though C-band penetration into dense mangrove canopy is limited compared with L-band; it responds primarily to the upper canopy and pneumatophore layer.
- Planet SuperDove: 3 m resolution, eight spectral bands including red-edge and NIR, near-daily revisit. Resolves narrow fringing stands that fall below Sentinel-2's pixel threshold. Commercial tasking adds cost but is the most practical option for monitoring stands less than 30 m wide.
What makes mangroves spectrally distinct, and what makes them difficult
Mangroves have unusually high leaf water content and a dense, multi-layered canopy structure that produces a strong NIR reflectance plateau and a steep red-edge slope. In Sentinel-2 band space, this places healthy mangrove canopy in a region clearly separated from dry terrestrial forest and from bare mudflat. The SWIR bands (1610 nm and 2190 nm) add sensitivity to canopy moisture and help distinguish mangrove from other wetland vegetation such as reed beds or saltmarsh.
The complication is that mangroves grow at the land-sea interface, and the spectral signal of the underlying tidal flat bleeds into the canopy signal wherever the stand is sparse or narrow. A pixel that is half water and half canopy reads as neither. This is not a processing artefact; it is physics. It means that fringing mangrove stands narrower than roughly two to three Sentinel-2 pixels (20 to 30 m) are systematically underestimated in extent and density. Planet SuperDove at 3 m reduces but does not eliminate this problem for the very narrowest stands.
Tidal phase is not a detail, it is the experiment design
Mangrove canopy reflectance changes measurably with tidal state. At high tide, standing water beneath the canopy absorbs NIR and darkens the apparent canopy signal. At low tide, exposed pneumatophores and mud contribute their own reflectance. An image acquired at mid-tide may produce a different NDVI or NDWI value than one acquired at low tide over the same stand, even if nothing biological has changed.
Consistent mapping therefore requires filtering the image archive by tidal phase at acquisition time, using co-located tide-gauge records or a global tidal model such as FES2014. For a Sentinel-2 time series this is practical: the 12-day revisit (single satellite) or 5-day revisit (two satellites) provides enough acquisitions per year to select low-tide scenes in most tropical locations. In regions with small tidal range (below roughly 0.5 m), tidal filtering matters less, but it should still be documented rather than assumed away.
Separating seasonal inundation from permanent loss
A mangrove stand that appears to shrink in a single dry-season image may simply be expressing the spectral effect of lower water tables and reduced canopy moisture, not actual dieback. Conversely, a stand stressed by hypersaline conditions during drought may look spectrally healthy until it collapses abruptly. Single-date classification is therefore unreliable for change detection.
The standard approach uses a dense Sentinel-2 time series, typically 12 to 24 tidal-filtered scenes per year, to build per-pixel phenological profiles. Permanent loss shows as a sustained step change in NIR reflectance and a corresponding increase in SWIR or bare-soil indices. Seasonal inundation shows as a recurring annual pattern that recovers. The Global Mangrove Watch (GMW) programme, produced from JAXA ALOS PALSAR and Sentinel-1 data, provides a published baseline at 25 m resolution with annual epochs from 1996 to 2020, which serves as a useful independent cross-check for any new analysis.
SAR backscatter under cloud: useful, but read the small print
Tropical mangrove regions frequently experience cloud cover that renders optical time series sparse for months at a time. Sentinel-1 C-band SAR acquires through cloud and provides consistent 12-day coverage (6-day with both satellites). In mangrove canopies, C-band double-bounce backscatter from the trunk-water interface is a known signal that strengthens when tidal inundation is present, and weakens when the stand is dry or structurally degraded.
The honest limit is penetration depth. C-band at 5.4 GHz interacts mainly with the upper metre or two of a dense mangrove canopy. It does not measure basal area or below-canopy structure reliably. L-band SAR (ALOS-2 PALSAR-2, or the forthcoming NISAR mission) penetrates further and is more sensitive to woody biomass, but L-band tasking is not routinely free. For canopy density estimation rather than simple extent mapping, SAR backscatter should be treated as a corroborating layer rather than the primary measurement.
Accuracy floors and what to report honestly
Published validation studies for regional mangrove maps typically report overall accuracies of 85 to 95 percent at the pixel level, but this headline figure conceals important variation. Accuracy is highest in wide, structurally intact stands and lowest at ecotone edges, in degraded or recovering patches, and in narrow fringing stands. Commission errors (labelling non-mangrove as mangrove) are common where aquatic vegetation or dense saltmarsh occupies similar spectral space.
Canopy density estimation, usually expressed as canopy cover fraction or a proxy such as NDVI, carries additional uncertainty. Published studies suggest that NDVI-derived canopy cover in mangroves has a root-mean-square error of roughly 10 to 20 percentage points against field measurements, depending on stand structure and tidal conditions at acquisition. Any delivered product should report these limits explicitly, alongside the tidal filtering criteria and the cloud-cover threshold applied to scene selection. Satellize applies this transparency as standard in its analytics outputs, including the crop-estimation work it runs for the Kingdom of Tonga, where similar phenological time-series methods are used.
For government clients commissioning baseline maps for carbon accounting or protected-area management, the key deliverable is not just the map but the documented uncertainty surface that accompanies it.
Putting the layers together: a practical workflow
A defensible mangrove extent and change product typically combines at least three inputs: a tidal-filtered Sentinel-2 time series for spectral classification, Sentinel-1 SAR backscatter for cloud-gap filling and structural corroboration, and the Global Mangrove Watch archive as a historical baseline. Planet SuperDove imagery is added where the target includes narrow fringing stands or where sub-annual change detection at fine scale is required.
Classification is usually performed with a supervised random forest or gradient-boosted model trained on field-verified reference points, with red-edge and SWIR bands as the highest-weight features. Change is detected by comparing per-pixel annual composites rather than individual scenes, which suppresses tidal and atmospheric noise. Outputs are delivered as annual GIS polygon layers with per-polygon confidence scores, a pixel-level change raster, and a summary table of net gain, net loss and stable area by administrative or management unit.
Typical figures
| Spatial resolution (optical) | 10 m (Sentinel-2 visible/NIR), 20 m (Sentinel-2 red-edge/SWIR), 30 m (Landsat 8/9), 3 m (Planet SuperDove) |
| Spatial resolution (SAR) | 10 m (Sentinel-1 IW mode, ground range) |
| Revisit cadence | 5 days (Sentinel-2, two satellites); 6 days (Sentinel-1, two satellites); 16 days (Landsat 8 or 9 individually); near-daily (Planet SuperDove) |
| Key spectral bands for mangrove discrimination | Red-edge (~705 nm, ~740 nm, ~783 nm), NIR (~842 nm), SWIR (~1610 nm, ~2190 nm); Sentinel-2 MSI bands 5, 6, 7, 8, 11, 12 |
| Minimum mappable stand width (Sentinel-2) | Approximately 20 to 30 m (2 to 3 pixels); narrower stands systematically underestimated |
| Minimum mappable stand width (Planet SuperDove) | Approximately 6 to 9 m in practice; sub-pixel mixing remains at edges |
| Archive depth | Sentinel-2 from 2015; Landsat from 1984 (Landsat 5 TM); Global Mangrove Watch baseline epochs from 1996 to 2020 |
| Tidal filtering requirement | Low-tide scenes preferred; tidal range above ~0.5 m makes filtering critical; FES2014 or co-located gauge records used for scene selection |
| Typical classification accuracy (published range) | 85 to 95% overall accuracy at pixel level; lower at ecotone edges and narrow stands |
| Canopy density estimation error (published range) | RMSE approximately 10 to 20 percentage points of canopy cover fraction against field measurements |
Analytics Satellize can run
| Annual mangrove extent map | Supervised random forest classification on tidal-filtered Sentinel-2 annual composites using red-edge, NIR and SWIR bands | GIS polygon layer (GeoPackage or Shapefile) with per-polygon confidence score and area statistics |
| Net change raster (gain, loss, stable) | Bi-temporal or multi-temporal comparison of annual composites; change confirmed across at least two consecutive epochs to suppress noise | Pixel-level change raster with summary table of net gain, net loss and stable area by management unit |
| Canopy density index time series | Per-pixel NDVI or EVI time series from tidal-filtered scenes; phenological decomposition to separate structural change from seasonal signal | Annual raster stack with per-pixel trend coefficient and significance flag |
| Cloud-gap-filled extent composite | Sentinel-1 C-band double-bounce backscatter fusion with optical classification to fill cloud-obscured periods | Monthly or seasonal composite raster with data-source flag per pixel |
| Narrow-stand fringe map | Planet SuperDove 3 m classification for target areas where Sentinel-2 pixel mixing is the binding constraint | High-resolution GIS layer with stand-width attribute and sub-stand density estimate |
| Uncertainty surface | Per-pixel posterior probability from classifier, combined with tidal-phase metadata and cloud-cover fraction at acquisition | Raster layer of classification confidence accompanying every extent product, with documented thresholds |
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.