Tropical forest phenology and deciduousness mapping
Tropical forests are not uniformly green year-round. Distinguishing evergreen, semi-deciduous and deciduous canopy types from multi-year EVI and NIR time series reveals forest function, drought sensitivity and land-use history that a single image cannot.
Sensors
- MODIS MOD13Q1 / MYD13Q1: 250 m resolution, 16-day composite (8-day when Terra and Aqua are combined via MOD13Q1 and MYD13Q1 interleaving). The workhorse for phenological time-series analysis; archive runs from 2000, giving more than two decades of seasonal cycles. Cloud masking is built into the product but humid-tropics scenes still carry significant contaminated epochs.
- VIIRS VNP13A1: 500 m, 16-day EVI and NDVI composites from Suomi-NPP and NOAA-20. Extends the MODIS-class record with a newer sensor and slightly different spectral response; useful for cross-validation and for post-2012 continuity when MODIS Terra degrades.
- Sentinel-2 MSI: 10 m (visible and NIR) to 20 m (red-edge, SWIR), revisit of five days at the equator with both satellites. Resolves phenological heterogeneity inside fragments too small for MODIS to separate. Cloud contamination in humid tropics can reduce usable observations to fewer than ten per year at a given pixel, making harmonic methods essential rather than optional.
- Landsat OLI (8 and 9): 30 m, 16-day revisit per satellite (eight days combined). Bridges the spatial gap between MODIS and Sentinel-2 with an archive stretching to 1984 for Landsat 5 TM, allowing phenological baseline comparison across decades. Surface reflectance Collection 2 products include cloud and cloud-shadow masks.
What leaf exchange actually signals
Deciduousness in tropical forests is not simply a response to cold, as it is in temperate zones. It is driven primarily by water-deficit stress during dry seasons, and the degree of leaf shedding scales with how severe and predictable that deficit is. A forest that drops 80 percent of its canopy in a four-month dry season behaves ecologically and hydrologically in a fundamentally different way from one that retains leaves year-round, even if both look 'tropical forest' on a coarse land-cover map.
That distinction matters for carbon accounting, fire-risk assessment, biodiversity surveys and water-balance modelling. Deciduous canopies expose soil, alter albedo and change the timing of litter inputs to the forest floor. Semi-deciduous stands, which shed leaves partially and asynchronously across species, are particularly hard to characterise from the ground but show a clear, measurable signal in multi-year EVI time series.
Why a single image fails and a time series does not
The spectral difference between a leafy deciduous canopy in the wet season and a nearby evergreen stand can be negligible. The difference becomes unambiguous only when you observe both through a full annual cycle. Evergreen canopies show low EVI amplitude across the year. Deciduous canopies show a strong seasonal trough aligned with the dry season. Semi-deciduous stands sit between those extremes, with intermediate amplitude and often a phase shift relative to fully deciduous neighbours.
MODIS MOD13Q1 8-day composites, stacked across five or more years, give roughly 230 observations per pixel per year in cloud-free conditions. In practice, over humid tropical regions, cloud contamination can reduce that to 40 to 80 usable observations annually. That is still enough to resolve the dominant annual harmonic if the reconstruction is done carefully. Sentinel-2 at five-day revisit sounds better, but a single humid-tropics site may yield fewer than 15 cloud-free images per year at 10 m, so the temporal density advantage of MODIS remains real.
Harmonic regression: reconstructing a signal from gaps
Harmonic regression fits a sum of sine and cosine terms to the observed reflectance or index values, treating the annual cycle as the fundamental frequency and adding a semi-annual harmonic to capture bimodal rainfall regimes. The fitted curve is continuous and can be evaluated at any date, regardless of whether an actual observation exists. The amplitude of the annual harmonic is the primary classifier: high amplitude indicates deciduous behaviour, low amplitude indicates evergreen.
The method is well-established in the published literature and is the basis of products such as the MODIS Land Cover Dynamics (MCD12Q2) phenology dataset, which reports greenup, maturity, senescence and dormancy dates globally. Applying the same harmonic framework to Sentinel-2 time series at 10 m requires careful quality filtering, since a single cloud-contaminated observation can distort the fit substantially. Iterative outlier rejection, weighting observations by their cloud-mask confidence score, and gap-filling using MODIS as a spatial prior are all documented approaches that improve stability.
One honest limit: harmonic regression assumes the phenological signal is periodic and stationary across the fitting window. In forests experiencing drought stress or land-cover change mid-record, the fitted amplitude conflates interannual variability with structural change. Fitting shorter windows (two to three years) and tracking amplitude trends over time is a better approach for dynamic landscapes, at the cost of noisier per-window estimates.
Classification: from amplitude to canopy type
The standard output is a three-class or continuous map: evergreen (annual EVI amplitude below roughly 0.1 in MODIS units, though thresholds vary by region), semi-deciduous (amplitude 0.1 to 0.2) and deciduous (amplitude above 0.2). These thresholds are not universal. Cerrado-adjacent dry forests in South America and miombo woodlands in southern Africa have different baseline EVI values and different cloud regimes, so regional calibration against field or high-resolution reference data is necessary.
Red-edge bands from Sentinel-2 (bands 5, 6 and 7, centred near 705, 740 and 783 nm) add sensitivity to chlorophyll content changes that precede visible leaf drop by several weeks. Including red-edge reflectance alongside broadband NIR improves the separation of semi-deciduous from evergreen classes, particularly in forests where canopy greenness is maintained visually but photosynthetic capacity is declining. This is a documented advantage of Sentinel-2 over MODIS for phenological discrimination at the sub-canopy-type level.
Practical limits the buyer should know
Cloud is the dominant constraint and it is not evenly distributed. The Congo Basin and the western Amazon can have fewer than 20 cloud-free MODIS observations per year in some pixels. Harmonic regression degrades gracefully as observation density falls, but below roughly 15 good observations per annual cycle the fitted amplitude becomes unreliable. Radar backscatter from Sentinel-1 can partially substitute, since SAR penetrates cloud, but the relationship between C-band backscatter and leaf-area phenology is weaker and more confounded by soil moisture than the optical EVI relationship.
Spatial resolution sets a hard floor on fragment detectability. At MODIS 250 m, a deciduous patch smaller than about 0.5 hectares is undetectable as a distinct class; it blends spectrally with neighbours. Sentinel-2 at 10 m lowers that floor to roughly 0.01 hectares, but only where the temporal record is dense enough to support the harmonic fit. In practice, the two sensors are complementary: MODIS for landscape-scale phenological mapping with high temporal confidence, Sentinel-2 for fragment-level classification where cloud allows.
Satellize runs phenological time-series pipelines on open Sentinel and MODIS archives, applying the same harmonic framework used in its Tonga crop-estimation programme to forest canopy contexts. The output is a per-pixel amplitude and phase map delivered as a GIS layer, with an accompanying uncertainty estimate derived from the residuals of the harmonic fit.
Typical figures
| Spatial resolution (landscape mapping) | 250 m (MODIS MOD13Q1), 500 m (VIIRS VNP13A1) |
| Spatial resolution (fragment-level mapping) | 10 m NIR / 20 m red-edge (Sentinel-2 MSI), 30 m (Landsat OLI) |
| Effective revisit (cloud-free observations, humid tropics) | 40 to 80 per year typical for MODIS; fewer than 15 per year typical for Sentinel-2 in persistently cloudy zones |
| Minimum fitting window for stable harmonic | 3 to 5 years recommended; minimum ~15 good observations per annual cycle |
| Key spectral bands | NIR (~841 nm MODIS band 2; Sentinel-2 band 8), red (~645 nm), red-edge (Sentinel-2 bands 5, 6, 7: 705, 740, 783 nm) |
| Minimum detectable deciduous patch | ~0.5 ha at MODIS 250 m; ~0.01 ha at Sentinel-2 10 m (subject to temporal record quality) |
| Archive depth | MODIS from 2000; Landsat from 1984 (TM); Sentinel-2 from 2015; VIIRS from 2012 |
| Primary index | EVI (Enhanced Vegetation Index); secondary: NIR reflectance, red-edge chlorophyll indices |
| Delivery formats | GeoTIFF per-pixel amplitude and phase maps, vector class polygons, CSV phenological metrics per polygon |
Analytics Satellize can run
| Annual EVI amplitude and phase map | Harmonic regression (first and second Fourier harmonics) fitted to multi-year MODIS or Sentinel-2 time series with iterative outlier rejection | GeoTIFF raster: amplitude, phase and R-squared per pixel; uncertainty layer from fit residuals |
| Canopy-type classification (evergreen / semi-deciduous / deciduous) | Amplitude thresholding with regional calibration; optional random-forest classifier trained on amplitude, phase, red-edge indices and terrain covariates | Vector polygon map with class probabilities; area statistics by administrative or protected-area boundary |
| Interannual amplitude trend | Linear trend fitted to per-year amplitude estimates across the archive; Mann-Kendall significance test | GIS layer showing pixels with significant greening or browning trends; summary report with confidence intervals |
| Phenological calendar (greenup, peak, senescence dates) | Derivative analysis of fitted harmonic curve; threshold-crossing dates extracted per pixel, following the MCD12Q2 published approach | Per-pixel date rasters (day-of-year) for four phenological transition points; multi-year mean and standard deviation |
| Cloud-gap-filled EVI composites | Harmonic curve used as spatial-temporal prior; Sentinel-2 observations gap-filled using MODIS-derived harmonic prediction where cloud masks flag missing data | Monthly or 16-day gap-filled EVI GeoTIFF stack at 10 m or 30 m for downstream analysis |
| Fragmented-landscape phenological mosaic | Sentinel-2 red-edge and NIR time series harmonically fitted at 20 m; class boundaries refined against MODIS landscape-scale amplitude map | High-resolution classification GeoTIFF with fragment-level deciduousness scores; patch-size distribution statistics |
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.