Coffee plantation canopy stress detection for crop-loan risk assessment
Red-edge and SWIR reflectance from Sentinel-2 and WorldView-3 detect chlorophyll loss and leaf-water deficit in coffee plantations weeks before visible defoliation, giving crop lenders early warning of collateral deterioration.
Sensors
- Sentinel-2 MSI: 10 m visible and NIR bands, 20 m red-edge bands B5 (705 nm), B6 (740 nm), B7 (783 nm) and SWIR bands B11/B12. Five-day revisit at the equator with both satellites. Red-edge bands are the primary chlorophyll-stress channel for coffee; SWIR tracks leaf water content. Free and open archive from 2015.
- WorldView-3 SWIR: Eight SWIR bands at 7.5 m resolution, covering 1195–2365 nm. Sensitive to canopy water content and early drought stress independently of chlorophyll signal. Tasked commercially; not a free archive. Published studies on Ethiopian coffee have used these bands to separate stress from shade-tree signal.
- Planet SuperDove: 3 m resolution, daily revisit in most tropical latitudes. Carries a red-edge band at 705 nm. Useful for resolving individual smallholder plots that are blurred in Sentinel-2's 20 m red-edge pixels. Commercial licence required; archive from 2021 for SuperDove specifically.
- Landsat 8/9 OLI: 30 m multispectral including SWIR1 (1565 nm) and SWIR2 (2200 nm). No dedicated red-edge band, so chlorophyll sensitivity is lower than Sentinel-2. Useful for long-baseline trend analysis given the archive depth to 1972 (Landsat 1) and free access. Sixteen-day revisit per satellite.
What the red-edge actually measures in a coffee canopy
Chlorophyll absorbs strongly in the red (around 670 nm) and reflects sharply in the NIR. The transition zone between those two regions, roughly 700–740 nm, is the red-edge. When chlorophyll concentration falls, that transition shifts toward shorter wavelengths and the reflectance in the 705–740 nm window rises relative to a healthy baseline. Sentinel-2 bands B5, B6 and B7 were designed to capture exactly this gradient. Indices derived from them, particularly the red-edge chlorophyll index (CIre) and the MERIS Terrestrial Chlorophyll Index (MTCI), correlate with leaf chlorophyll content measured in the laboratory across a wide range of broadleaf crops.
Coffee is physiologically unusual in that stress often begins at the root zone, from drought or fungal infection, and propagates upward before leaves show visible yellowing. The spectral signal therefore precedes the agronomic symptom by two to six weeks in published field studies, depending on stress type and severity. For a lender assessing collateral on a seasonal crop loan, that lead time is the entire value of the observation.
The shade-canopy problem, stated plainly
Most Arabica coffee grown at altitude, particularly in Ethiopia's Sidama and Yirgacheffe zones and Colombia's Andean smallholder plots, is cultivated under a partial or full shade canopy of Erythrina, Grevillea or native forest trees. At Sentinel-2's 20 m red-edge pixel, a mixed signal from shade trees and coffee shrubs is unavoidable in fragmented landscapes. The shade trees may be healthy while the coffee underneath is stressed, or vice versa. Spectral unmixing algorithms can partially separate the two signals if the shade-tree species are spectrally distinct and their fractional cover is mapped, but this requires calibration data and introduces uncertainty.
WorldView-3 at 7.5 m SWIR resolution reduces the mixing problem for larger plots. Planet SuperDove at 3 m reduces it further for individual rows. Neither eliminates it entirely where shade cover exceeds roughly 60 per cent. Any stress index delivered to a lender should carry an explicit confidence flag tied to estimated shade-canopy fraction in each polygon.
Smallholder fragmentation and the pixel-purity floor
Ethiopian coffee smallholdings average well under one hectare. A single Sentinel-2 20 m pixel covers 400 square metres; a plot of that size contains perhaps four pixels, and boundary pixels mix coffee with adjacent land cover. The practical consequence is that Sentinel-2 red-edge indices are more reliable for cooperative-aggregated loan books, where the effective unit of analysis is a village-level cluster of several hundred hectares, than for individual smallholder loan assessment.
Colombian coffee landscapes present a similar fragmentation challenge on steep Andean slopes, compounded by cloud cover. The Andean coffee belt receives 1800–3000 mm of rainfall annually and cloud obscuration can run at 60–80 per cent of acquisition opportunities in the main growing season. A five-day Sentinel-2 revisit does not guarantee five-day clear-sky observations. Realistic clear-sky compositing intervals in these landscapes are often 20–40 days, which narrows but does not eliminate the stress-detection lead time advantage.
From index to default-risk signal: the analytical chain
A stress-detection workflow for crop-loan risk begins with plantation boundary polygons, ideally from a field-verified or high-resolution-derived map. Per-polygon time series of CIre or MTCI are extracted from all cloud-free Sentinel-2 acquisitions across the growing season. A z-score relative to the same phenological stage in prior years flags anomalous canopy condition. The flag is then weighted by the loan book's exposure to that polygon or cluster, producing a portfolio heat map rather than a single-farm alert.
The SWIR dimension adds a separate line of evidence. Sentinel-2 B11 (1610 nm) and B12 (2190 nm) respond to leaf water content through liquid-water absorption features. The normalised difference water index (NDWI, using NIR and SWIR) declines measurably before chlorophyll loss is detectable in the red-edge, making it a useful early-warning complement rather than a substitute. Combining both signals reduces false-positive rates caused by temporary cloud shadow or sensor artefacts.
What the archive can and cannot tell a lender
Sentinel-2's archive from 2015 provides roughly a decade of seasonal baselines for Ethiopian and Colombian coffee regions. That is long enough to characterise inter-annual variability driven by El Niño and La Niña cycles, which are the dominant drivers of regional drought stress in both landscapes. A loan underwriter can therefore ask not just whether this season's canopy is stressed, but whether it is stressed relative to a drought year that caused documented defaults.
What the archive cannot tell you is the cause of the stress signal. Chlorophyll loss and water deficit are symptoms shared by drought, coffee leaf rust (Hemileia vastatrix), coffee wilt disease (Gibberella xylarioides), and nutrient deficiency. Distinguishing among these requires either very high-resolution imagery with texture analysis, field sampling, or ancillary disease-monitoring data. Satellize incorporates ancillary agronomic and weather layers into its plantation analytics, as it does in the Tonga crop-estimation programme, but the spectral data alone does not resolve disease aetiology. Lenders should treat the stress signal as a trigger for field verification, not a standalone loss estimate.
Delivering the analysis to a credit portfolio
The practical deliverable for a crop-finance institution is a ranked list of loan exposures by canopy-stress severity, updated at each clear-sky composite interval. Each entry carries the plantation polygon identifier, the current-season CIre anomaly score, the SWIR water-stress index, estimated shade-canopy fraction as a confidence qualifier, cloud-cover fraction for the observation window, and a traffic-light flag. GIS layers in GeoJSON or GeoPackage format integrate directly into most loan-management systems that carry geographic collateral fields.
For parametric crop-loan products, the stress index can be written into the loan contract as a monitoring trigger at a defined threshold, removing the need for field adjustment entirely. The index threshold and the observation window need to be agreed at origination and verified against historical loss data. That calibration step is where the value of a decade-long Sentinel-2 archive is most directly monetised.
Typical figures
| Primary spatial resolution (red-edge) | 20 m (Sentinel-2 B5/B6/B7); 3 m (Planet SuperDove red-edge); 7.5 m (WorldView-3 SWIR) |
| Revisit interval | 5 days (Sentinel-2, equatorial); daily (Planet SuperDove); tasked on demand (WorldView-3) |
| Effective clear-sky revisit (Andean/Ethiopian highlands) | 20–40 days typical in peak cloud season; shorter in dry season |
| Key spectral bands | Red-edge: 705 nm, 740 nm, 783 nm (Sentinel-2 B5/B6/B7); SWIR: 1610 nm, 2190 nm (Sentinel-2 B11/B12); SWIR: 1195–2365 nm (WorldView-3) |
| Primary stress indices | Red-edge chlorophyll index (CIre), MERIS Terrestrial Chlorophyll Index (MTCI), Normalised Difference Water Index (NDWI) |
| Minimum reliable polygon size (Sentinel-2) | Approximately 1 ha for red-edge; smaller polygons require Planet or WorldView-3 |
| Lead time before visible defoliation | 2–6 weeks in published field studies, stress-type dependent |
| Archive depth | Sentinel-2: from 2015; Landsat: from 1972 (30 m, no red-edge); Planet SuperDove: from 2021 |
| Delivery formats | GeoJSON, GeoPackage, GeoTIFF, tabular CSV with polygon identifiers |
| Update latency (Sentinel-2 to deliverable) | Typically 24–72 hours after cloud-free acquisition, depending on processing pipeline |
Analytics Satellize can run
| Per-polygon canopy stress score | Red-edge chlorophyll index (CIre) and MTCI derived from Sentinel-2 B5/B6/B7; z-score anomaly relative to same phenological stage in prior seasons | Ranked GIS layer and CSV table, updated per clear-sky composite |
| Leaf water deficit index | NDWI from Sentinel-2 NIR (B8) and SWIR (B11); threshold-based alert when index falls below plantation-specific baseline | Time-series chart per loan polygon; alert flag in portfolio dashboard |
| Shade-canopy fraction map | Spectral unmixing or supervised classification using high-resolution optical imagery to estimate shade-tree cover fraction per plantation polygon; used as confidence weight on stress scores | GeoTIFF confidence layer appended to stress score deliverable |
| Seasonal stress trajectory report | Phenology-normalised time series of CIre and NDWI across the growing season, benchmarked against El Niño and La Niña baseline years from Sentinel-2 archive | PDF seasonal report per cooperative or loan-book cluster, with historical percentile ranking |
| Portfolio heat map for credit review | Stress scores weighted by outstanding loan exposure per polygon; aggregated to district or cooperative level | Interactive GIS layer or static PDF map for credit-committee review |
| Parametric trigger monitoring | Automated comparison of composite-period stress index against contractually defined threshold; binary pass/fail output with supporting evidence package | Trigger verification report with underlying imagery, index value, cloud-cover fraction and observation date |
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.