Forest phenology and seasonality mapping
Dense satellite time-series reveal when forests flush, peak and senesce, exposing differences between forest types and flagging drought stress weeks before visible dieback. This page explains the sensors, methods and honest limits of spaceborne phenology mapping.
Sensors
- MODIS Terra/Aqua: 250 m (bands 1–2) to 500 m resolution; daily overpass per platform, combined giving twice-daily coverage. The standard MOD13 and MYD13 vegetation-index products composite to 8-day and 16-day intervals, smoothing cloud gaps sufficiently for phenology curve-fitting at continental scale. Archive runs from 2000, giving more than two decades of seasonal baselines.
- VIIRS SNPP / NOAA-20: 375 m I-band resolution for vegetation indices; daily revisit. The VNP13 and VJ113 products follow the MODIS heritage methodology and extend the time series forward from 2012 (SNPP) and 2018 (NOAA-20), important for trend continuity as MODIS instruments age.
- Sentinel-2 MSI (2A and 2B combined): 10 m resolution in the red, green, blue and near-infrared bands; 20 m in the red-edge bands (B5, B6, B7) that are particularly sensitive to chlorophyll content changes during flush and senescence. Combined 2A/2B revisit is approximately 5 days at the equator and shorter at higher latitudes. Cloud remains the limiting factor: in persistently cloudy tropical regions, usable observations may number fewer than 20 per year.
- Landsat 8/9 OLI: 30 m resolution across visible, near-infrared and shortwave-infrared bands. The 16-day single-satellite revisit (8 days when both are combined) is marginal for capturing rapid leaf-flush events but adequate for end-of-season and growing-season-length metrics. The archive back to 1972 (earlier sensors) and the consistent OLI calibration from 2013 make Landsat indispensable for multi-decade phenology trend analysis.
What the seasonal curve actually measures
Forest phenology, in remote-sensing terms, is the shape of a vegetation-index time series plotted across a calendar year. The normalised difference vegetation index (NDVI) or the enhanced vegetation index (EVI) rises as leaves emerge, plateaus during peak greenness, then falls as chlorophyll degrades in autumn or during a dry-season dormancy. The inflection points on that curve correspond to biologically meaningful dates: start of season (SOS), end of season (EOS), peak greenness and the growing season length (GSL) derived from them.
The red-edge bands on Sentinel-2 (centred near 705 nm, 740 nm and 783 nm) add sensitivity that broadband NDVI lacks. Chlorophyll begins to decline measurably in the red-edge reflectance before it is visible to the eye or detectable in standard NDVI, giving roughly a one-to-two-week earlier warning of senescence onset. That lead time matters when the goal is detecting drought-induced stress rather than simply recording it after the fact.
Deciduous, semi-deciduous, evergreen: the signal differences that enable classification
Deciduous broadleaf forests produce a high-amplitude, clearly bounded seasonal cycle. Tropical evergreen forests produce a low-amplitude cycle modulated by dry-season intensity rather than temperature. Semi-deciduous forests sit between the two, with partial leaf exchange whose timing varies by species and microclimate. These differences in curve shape, amplitude and phase are the basis on which phenology mapping supports forest-type discrimination, complementing single-date classification approaches that cannot resolve them.
In practice, the separation is cleaner at coarser resolutions (MODIS, VIIRS) where mixed pixels average out fine-scale heterogeneity and produce smoother curves. At Sentinel-2 resolution, individual tree crowns or small stands produce noisier time series, requiring more sophisticated gap-filling and smoothing before curve-fitting is reliable. Neither scale is universally superior; the choice depends on whether you need continental coverage or stand-level detail.
Extracting metrics: harmonic regression versus threshold methods
Two families of method dominate operational phenology extraction. Threshold-based approaches, used in the MODIS Land Cover Dynamics product (MCD12Q2), define SOS and EOS as the dates when a smoothed vegetation-index curve crosses a fixed or relative amplitude threshold. They are computationally simple and interpretable, but the threshold value is somewhat arbitrary and can shift apparent phenology dates if background reflectance changes between years.
Harmonic regression fits a sum of sine and cosine functions to the annual time series. It handles irregular sampling and cloud-gap data more gracefully and produces continuous amplitude and phase parameters that can themselves be used as classification features. The published TIMESAT software package implements both approaches and is widely used in the literature. A practical limitation of both methods is that they require a minimum of roughly 20 to 30 cloud-free observations per year to produce reliable metrics. In humid tropical regions, this is often not achievable with a single sensor, making sensor fusion (combining Sentinel-2 and Landsat, for instance) a necessity rather than a refinement.
Whichever method is used, the output phenology metrics carry uncertainty that should be reported alongside the metric itself. A SOS date derived from a sparse, partially cloud-contaminated time series may carry an uncertainty of plus or minus two to three weeks. Presenting a single date without that range misleads the user.
Drought stress detection before the canopy shows it
One of the more practically valuable applications of phenology time series is anomaly detection: comparing the current year's vegetation-index trajectory against the multi-year baseline for the same pixel. A canopy that reaches peak greenness two weeks later than its historical median, or that begins senescence three weeks early, is exhibiting a measurable stress signal even if aerial inspection would not yet flag anything unusual.
MODIS's archive depth (from 2000) is critical here. Detecting a meaningful anomaly requires knowing what normal looks like across a range of years that includes both wet and dry conditions. A single anomalous year in a short record is ambiguous; the same anomaly against a 20-year baseline, with its inter-annual variance characterised, is informative. The 2005 and 2010 Amazon droughts, for example, produced detectable EVI anomalies in MODIS data that preceded large-scale canopy mortality by weeks, a finding documented in published literature using exactly this approach.
The limit is spatial resolution. At 250 m to 500 m, MODIS cannot resolve stress in individual stands smaller than a few hectares. Sentinel-2 can, but its shorter archive (from 2017 for consistent global coverage) means the baseline is thinner. Combining the two, using MODIS for the long baseline and Sentinel-2 for spatial detail, is the approach that makes operational sense.
Phenology as an input to carbon flux models
Growing season length is a primary driver of gross primary production (GPP) in temperate and boreal forests. A shift of ten days in SOS or EOS, sustained over years, changes annual carbon uptake estimates by amounts that are significant relative to national greenhouse gas inventory uncertainties. This is why phenology metrics feed directly into carbon flux models such as MODIS GPP (MOD17) and into the terrestrial biosphere models used in IPCC assessments.
Satellize applies harmonic and threshold phenology extraction to Sentinel-2 and MODIS/VIIRS stacks as part of its analytics work on open constellations, with the Tonga crop-estimation programme providing a practical test bed for the same time-series methods applied to agricultural rather than forest canopies. The underlying approach transfers directly.
For national forest inventories and REDD+ programmes, phenology layers add a temporal dimension to what are otherwise static land-cover maps. A pixel classified as tropical moist forest in a single-date map may behave, phenologically, more like a degraded or secondary forest. The time series reveals what the snapshot conceals.
What this method cannot do
Persistent cloud cover in equatorial regions can reduce usable Sentinel-2 observations to fewer than ten per year over some areas, making reliable curve-fitting impossible without sensor fusion or gap-filling from coarser sensors. Even with fusion, the uncertainty in derived metrics grows substantially.
Phenology mapping detects canopy-level changes in reflectance. It does not directly measure species composition, stem density or below-canopy structure. A monoculture plantation and a species-rich natural forest can produce similar NDVI curves if their canopy closure and leaf-area index are comparable. Distinguishing them requires additional inputs, whether hyperspectral data, lidar or field calibration.
Finally, phenology metrics are sensitive to the smoothing and gap-filling algorithm chosen. Different software implementations applied to the same raw time series can produce SOS dates that differ by one to two weeks. Any multi-site or multi-year comparison should use a consistent processing chain throughout.
Typical figures
| Spatial resolution (phenology metrics) | 250–500 m (MODIS/VIIRS); 10–20 m (Sentinel-2 MSI); 30 m (Landsat 8/9 OLI) |
| Temporal compositing interval | 8-day and 16-day (MODIS standard products); 5-day effective revisit (Sentinel-2A+2B combined at equator) |
| Minimum usable observations per year | 20–30 cloud-free scenes for reliable harmonic or threshold curve-fitting |
| Key spectral bands | Red (660–670 nm), NIR (841–876 nm) for NDVI/EVI; red-edge (705, 740, 783 nm) on Sentinel-2 for early stress detection |
| Archive depth | MODIS: from 2000; VIIRS SNPP: from 2012; Sentinel-2: consistent global from 2017; Landsat: from 1972 (OLI from 2013) |
| Typical SOS/EOS date uncertainty | ±1–3 weeks depending on cloud frequency and observation density |
| Minimum detectable anomaly | Phenology shifts of approximately 10–14 days detectable against a multi-year baseline with adequate observation density |
| Coverage | Global; tropical cloud-affected regions require multi-sensor fusion for adequate annual observation count |
| Standard output formats | GeoTIFF raster layers (per metric per year); CSV phenology-date tables; time-series NetCDF stacks |
| Latency (operational products) | MODIS/VIIRS near-real-time composites available within 1–2 days of acquisition; Sentinel-2 L2A typically within 3–5 days of acquisition via Copernicus Data Space |
Analytics Satellize can run
| Annual phenology metric maps (SOS, EOS, GSL, peak NDVI date) | Threshold-based extraction (TIMESAT-style) or harmonic regression applied to gap-filled MODIS MOD13 or Sentinel-2 NDVI/EVI stacks | Annual GeoTIFF raster layers per metric, with per-pixel uncertainty estimates, delivered as a GIS-ready package |
| Multi-year phenology trend analysis | Linear trend fitting on per-pixel SOS/EOS time series across the MODIS archive (2000 to present); Mann-Kendall significance testing | Trend magnitude and significance rasters; summary report identifying areas of statistically significant phenological shift |
| Current-year phenology anomaly detection | Z-score or percentile comparison of current-year vegetation-index trajectory against the per-pixel multi-year baseline; flagging of pixels exceeding defined anomaly thresholds | Anomaly alert layer updated at each compositing interval; optional automated alert feed for areas of interest |
| Forest-type discrimination using phenology curve shape | Harmonic amplitude and phase parameters used as classification features in a supervised random-forest or support-vector-machine classifier, trained on reference forest-type data | Forest-type probability map distinguishing evergreen, semi-deciduous and deciduous classes; confusion matrix and accuracy assessment |
| Drought stress early-warning layer | Comparison of current-season EVI/red-edge trajectory against baseline; detection of delayed SOS or early EOS relative to historical median | Stress-flag raster updated seasonally; tabular report of affected area by administrative unit or management zone |
| Growing season length input for GPP modelling | Per-pixel GSL derived from SOS/EOS metrics; formatted for ingestion into MODIS MOD17-compatible or custom terrestrial biosphere model inputs | GSL raster layer with metadata conforming to standard carbon-modelling input specifications; documentation of processing chain for audit purposes |
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.