Banana plantation health decline mapping
Progressive banana plantation decline, whether from Fusarium wilt, Sigatoka or other stresses, shows spectral signatures weeks before visual symptoms are obvious. Time-series red-edge and shortwave-infrared indices from Sentinel-2 and PlanetScope can map patch boundaries and guide targeted ground scouting.
Sensors
- Sentinel-2 MSI: 10 m resolution in visible and near-infrared bands; 20 m in red-edge (B5, B6, B7) and shortwave-infrared bands. Five-day revisit at the equator with both satellites. Red-edge bands are particularly sensitive to chlorophyll content and early stress before visible yellowing. Free and open archive from 2015.
- PlanetScope: 3 m resolution, daily revisit over most tropical latitudes. Eight-band SuperDove instruments include red-edge and near-infrared. High spatial resolution resolves individual plant rows and small decline patches (roughly 10 plants or more), though the archive depth and band calibration consistency vary by acquisition date.
- WorldView-3: 0.31 m panchromatic, 1.24 m multispectral, 3.7 m shortwave-infrared (eight SWIR bands). Tasked commercially. SWIR bands can detect leaf water content changes and, in principle, fungal-induced cell collapse, but tasking cost limits use to high-value confirmation of flagged patches rather than routine monitoring.
- Airborne VNIR hyperspectral: Typically 1–5 m spatial resolution with 100-plus contiguous spectral bands from 400 to 1000 nm. Published studies have used airborne hyperspectral data to discriminate Fusarium wilt symptom classes from healthy banana canopy with high accuracy, but coverage is limited to campaign areas and is expensive per hectare.
What a dying banana plant actually looks like from orbit
Banana pseudostems are mostly water. A healthy Cavendish canopy reflects strongly in the near-infrared and absorbs heavily in the red, producing a high NDVI. When Fusarium oxysporum f.sp. cubense (Foc) or Black Sigatoka invades, chlorophyll degrades from the inside out. The red-edge reflectance slope, the steep rise in reflectance between roughly 700 and 740 nm, flattens measurably before the leaf turns visibly yellow. This is the detection window that satellite red-edge indices are designed to exploit.
Sentinel-2's B5 (705 nm), B6 (740 nm) and B7 (783 nm) bands allow computation of the Red-Edge Chlorophyll Index (CIre) and the MERIS Terrestrial Chlorophyll Index (MTCI). Both are more sensitive to early chlorophyll loss than NDVI, which saturates at high canopy density. Published plantation-monitoring studies have used CIre time series to detect stress patches two to four weeks before they become obvious to ground scouts, though this lead time depends heavily on cloud-free imagery availability and the speed of pathogen progression.
Building a time series that is actually usable in the tropics
Cloud cover is the central operational problem. Banana-growing regions in Central America, Southeast Asia and West Africa routinely see cloud cover exceeding 70 percent of days. A single Sentinel-2 image is often useless. The practical answer is a dense time-series stack, typically 12 to 24 months of all available acquisitions, from which cloud-free composites are constructed using pixel-quality flags and median or percentile compositing. This approach, well established in the Copernicus land-monitoring literature, smooths atmospheric noise and reveals the underlying phenological trajectory of each pixel.
PlanetScope's daily cadence helps fill cloud gaps but introduces its own complications. Radiometric consistency across the SuperDove fleet is imperfect, and surface reflectance conversion quality varies. Cross-calibrating PlanetScope against Sentinel-2 on cloud-free overlap dates is standard practice before fusing the two datasets into a single time series. The fusion gives roughly 3 m spatial detail on the days PlanetScope is clear and 10 to 20 m detail otherwise, which is usually sufficient to resolve plantation blocks of commercial scale.
From index anomaly to decline patch boundary
The analytic workflow has three stages. First, a baseline is established from the earliest clean imagery, or from a historical archive period before any known outbreak. Second, per-pixel z-scores or cumulative sum (CUSUM) statistics are computed on the CIre or MTCI time series to flag pixels where the index has declined beyond a threshold relative to the baseline. Third, flagged pixels are spatially aggregated into candidate decline patches, filtered by minimum area (typically 0.1 to 0.5 hectares to suppress noise) and ranked by severity and rate of spread.
The output is not a disease map in a clinical sense. It is a ranked list of spatial anomalies that warrant ground investigation. This distinction matters. Satellite indices cannot distinguish Fusarium wilt from waterlogging stress, nematode damage, or simple senescence without additional context. The satellite layer narrows the search space; the pathologist on the ground makes the diagnosis. Honest use of the method treats it as a scouting prioritisation tool, not a confirmed disease inventory.
Resolution limits and what they mean for containment planning
At 10 m, a single Sentinel-2 pixel covers roughly 100 square metres, which is approximately 15 to 25 banana plants depending on spacing. An emerging Fusarium patch of five or six plants will not register as a distinct anomaly at this resolution; it will only appear when it has grown to affect most of a pixel. PlanetScope at 3 m reduces this threshold considerably, but even at 3 m a patch must comprise several plants to be reliably detected above noise.
This resolution floor has a direct implication for containment strategy. Satellite monitoring is most useful for detecting mid-stage spread across plantation blocks, not for catching the very first infected plant. Early-stage detection still requires systematic ground scouting, ideally informed by the satellite layer so scouts concentrate on high-risk zones. For Tropical Race 4 (TR4) outbreaks, where containment depends on acting before the pathogen establishes in soil, this lag is a genuine limitation that buyers should factor into their surveillance design.
Hyperspectral evidence and what it adds
Airborne VNIR hyperspectral campaigns have demonstrated that Fusarium-infected banana plants show characteristic absorption features near 550 nm and 680 nm that differ statistically from drought-stressed or Sigatoka-affected plants. With 100-plus narrow bands, a trained classifier can assign symptom classes with reported overall accuracies in the range of 80 to 90 percent in published studies, though these figures come from controlled campaign conditions and degrade in operational settings with mixed stress types.
Spaceborne hyperspectral is not yet routine for this application. PRISMA (ASI) and DESIS (DLR) offer hyperspectral data from orbit at roughly 30 m resolution, which is too coarse to resolve individual plant stress in most commercial plantations. The next generation of commercial hyperspectral satellites may change this calculus, but for now airborne hyperspectral is best treated as a targeted diagnostic tool for confirmed outbreak zones rather than a monitoring backbone.
Putting it into practice
A practical monitoring programme combines Sentinel-2 time-series analysis for plantation-wide surveillance, PlanetScope for spatial detail when cloud permits, and WorldView-3 or airborne hyperspectral tasked reactively over flagged anomaly zones. Alert cadence is typically fortnightly, constrained by cloud-free Sentinel-2 availability rather than satellite revisit. Latency from image acquisition to delivered GIS layer is usually one to three days for automated pipelines.
Satellize runs this class of vegetation-index time-series analytics on open constellations, with commercial tasking added on client licence. The Tonga crop-estimation programme is one example of the organisation applying dense time-series methods in a small-island agricultural context where ground access is limited. For banana producers or national plant-health agencies considering a monitoring programme, the logical first step is a retrospective analysis over a known historical outbreak area to establish what the signal actually looked like in your specific plantation type and environment before committing to operational deployment.
Typical figures
| Spatial resolution (routine monitoring) | 10 m (Sentinel-2 visible/NIR), 20 m (Sentinel-2 red-edge/SWIR), 3 m (PlanetScope) |
| Spatial resolution (targeted confirmation) | 1.24 m multispectral (WorldView-3); 1–5 m typical for airborne hyperspectral campaigns |
| Revisit cadence | 5 days (Sentinel-2 dual satellite); daily (PlanetScope, cloud-permitting); on-demand tasking (WorldView-3) |
| Key spectral bands | Red-edge B5/B6/B7 (Sentinel-2, 705–783 nm); red-edge band (PlanetScope SuperDove, ~705 nm); SWIR 1–8 (WorldView-3, 1195–2365 nm); 400–1000 nm continuous (airborne VNIR hyperspectral) |
| Minimum detectable decline patch | Approximately 0.1–0.5 ha at Sentinel-2 resolution; smaller patches detectable with PlanetScope fusion, subject to cloud-free availability |
| Typical time-series archive depth | Sentinel-2 from 2015; PlanetScope from approximately 2016 (coverage and band consistency variable by region) |
| Cloud cover constraint | Major operational limit in humid tropics; cloud-free compositing over 12–24 month stacks is standard mitigation |
| Alert latency | 1–3 days from image acquisition to GIS alert layer for automated pipelines |
| Delivery formats | GeoTIFF anomaly maps, vector patch boundaries (GeoJSON/Shapefile), ranked scouting priority report (PDF or dashboard) |
| Coverage | Global for Sentinel-2 and PlanetScope; WorldView-3 and airborne campaigns are area-specific and tasked |
Analytics Satellize can run
| Baseline canopy health map | Median composite of Sentinel-2 red-edge indices (CIre, MTCI) over a defined reference period, per published Copernicus land-monitoring compositing methods | GeoTIFF raster and plantation-block summary table (PDF or CSV) |
| Fortnightly decline anomaly alert | CUSUM or z-score change detection on dense CIre/MTCI time series relative to baseline; minimum-area filtering to suppress noise | Vector GeoJSON of flagged patches ranked by severity score, delivered to client GIS or dashboard |
| Patch growth rate and spread direction analysis | Sequential patch-boundary comparison across alert epochs; centroid tracking and area-change calculation | Fortnightly spread report with mapped patch trajectories and estimated expansion rate (ha per week) |
| Ground-scouting priority layer | Ranked overlay of anomaly severity, patch size and proximity to healthy blocks; standard spatial prioritisation logic | Printable field map and mobile-compatible GIS layer for scout routing |
| PlanetScope-Sentinel-2 fused decline map | Cross-calibrated fusion of PlanetScope daily imagery with Sentinel-2 surface reflectance on cloud-free overlap dates; spatial sharpening to 3 m | High-resolution GeoTIFF anomaly map for confirmed outbreak blocks |
| Historical outbreak retrospective | Archive time-series analysis over a known past outbreak period to characterise the index signature specific to the client's plantation type and environment | Calibration report establishing detection thresholds and expected lead time for the client's operational context |
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.