Crop insurance loss adjustment using post-event multispectral change detection
Pre- and post-event multispectral imagery quantifies crop damage extent and severity for indemnity loss adjustment, cutting settlement time and adjuster cost while providing an auditable spatial record.
Sensors
- Sentinel-2 MSI: 10 m resolution in red, green, blue and near-infrared bands; 20 m in red-edge and shortwave infrared. Five-day revisit at mid-latitudes with both satellites. Free archive from 2015. The workhorse for field-scale NDVI change detection on parcels above roughly 0.5 ha.
- Planet SuperDove: 3 m resolution, eight spectral bands including two red-edge channels. Daily revisit over most agricultural regions. Enables detection of damage on parcels too small for Sentinel-2 to resolve cleanly, though the archive depth is shallower and data access requires a commercial licence.
- Landsat 8/9 OLI: 30 m resolution, free archive from 1982. Sixteen-day revisit per satellite, eight days combined. Coarser than Sentinel-2 but the long archive provides multi-year phenological baselines that help distinguish peril damage from normal inter-annual variation in canopy greenness.
- Planet SkySat: 0.5 m native resolution, tasked on demand. Used for spot verification of disputed parcels rather than area-wide loss mapping. At this scale, lodging, hail bruising and flood scour are visually interpretable without spectral analysis.
Why near-infrared tells you what the visible spectrum cannot
Healthy green vegetation reflects strongly in the near-infrared (NIR, roughly 750–900 nm) because intact mesophyll cell walls scatter NIR photons back toward the sensor. Damaged or dead plant tissue loses that internal structure. NIR reflectance drops sharply, while red reflectance rises as chlorophyll degrades. The normalised difference vegetation index (NDVI = (NIR − Red) / (NIR + Red)) captures this contrast in a single dimensionless number, typically ranging from around 0.7–0.9 for a dense healthy canopy to below 0.2 for bare soil or fully destroyed crop.
A post-event image alone is not enough. Loss adjustment requires a comparison: the same field, the same crop, at the same phenological stage, in a year without the peril. That comparison can come from a pre-event image taken days earlier, from the same calendar window in prior seasons, or from spatially adjacent reference fields of the same variety that escaped the event. Without a credible reference, a low NDVI reading is ambiguous. It could mean hail damage, drought stress, disease, or a field that was simply harvested before the adjuster arrived.
The harvest problem: damage that looks identical to success
The single most common source of dispute in satellite-based loss adjustment is the harvested field. A combine harvester removes the canopy in hours. The post-event NDVI of a freshly harvested field is indistinguishable from a field destroyed by fire or flood if you look only at the post-event image. Resolving the ambiguity requires either a pre-event image captured while the crop was still standing, or a time-series dense enough to distinguish the abrupt one-day transition of harvest from the more gradual decline associated with most weather perils.
Sentinel-2's five-day revisit and Planet's daily cadence both support this kind of dense time-series analysis. In practice, cloud cover is the limiting factor. In humid tropical or monsoon climates, finding a cloud-free image within a week of an event can be impossible with optical sensors alone. SAR (covered on a sibling page) partially fills that gap, but SAR-based canopy damage detection is a different and less mature method. Honest programmes plan for optical latency of one to three weeks in cloudy seasons and communicate that uncertainty to claims handlers before an event, not after.
Minimum field size and the resolution floor
Spatial resolution sets a hard lower bound on what can be adjusted remotely. A pixel in Sentinel-2's NIR band is 10 m × 10 m, or 0.01 ha. But a field needs to occupy many pixels before its mean NDVI is statistically reliable. Mixed pixels at field edges contaminate readings. A practical rule of thumb, supported by published validation work in European agricultural contexts, is that Sentinel-2 gives reliable parcel-level NDVI estimates for fields above approximately 0.5 ha. Below that, Planet SuperDove at 3 m resolution extends coverage to parcels of roughly 0.05 ha, though at higher data cost.
For smallholder agriculture, particularly in sub-Saharan Africa or South and Southeast Asia where individual plots can be as small as 0.1–0.2 ha, even Planet struggles. SkySat at 0.5 m can resolve individual rows within a plot, but tasking costs make systematic coverage of large loss events impractical at that resolution. The honest answer for smallholder portfolios is that satellite loss adjustment works best when combined with a stratified field-visit programme: satellites identify which villages or zones were affected and at what severity, field adjusters then sample within those zones rather than visiting every plot.
Converting an NDVI anomaly into an indemnity figure
An NDVI change map tells you where the canopy is damaged. It does not automatically tell you what fraction of yield was lost, which is the number an insurer needs. The conversion from index anomaly to yield loss is crop-specific, phenology-specific, and peril-specific. A 0.3-point NDVI drop in wheat at grain-fill has different yield implications than the same drop at tillering. Hail damage to a maize canopy can leave NDVI relatively intact while destroying the ear entirely.
Calibrating that conversion requires ground-truth data: harvested yield from damaged and undamaged reference plots, matched to the satellite signal. Satellize's crop-estimation programme in the Kingdom of Tonga has generated exactly this kind of index-to-yield validation dataset for Pacific island crop types, demonstrating that the calibration methodology is transferable to new geographies when sufficient ground data are collected at programme inception. Insurers entering a new market should budget for one to two seasons of parallel field measurement before relying on satellite-derived loss figures for settlement without adjuster confirmation.
What the audit trail is actually worth
One underappreciated advantage of satellite loss adjustment is the permanence of the record. A field adjuster's visit produces a report and perhaps a few photographs. A satellite archive produces a dated, georeferenced, radiometrically calibrated image that can be re-analysed years later if a claim is disputed. Sentinel-2 data are archived by ESA and the Copernicus Data Space; Landsat data are archived by USGS. Both are freely accessible. That archive is as useful for fraud detection as it is for legitimate loss verification: a claimant who reports hail damage on a date when satellite imagery shows an undamaged canopy has a problem.
Delivery formats matter for integration into claims workflows. GIS polygon layers with per-parcel NDVI anomaly scores can be ingested directly into claims management systems. Raster damage-probability maps suit portfolio-level triage. A simple ranked list of affected policy numbers, sorted by estimated damage severity, is often what a claims manager actually needs on day one after a catastrophic event. The analytic product should match the workflow, not the other way around.
Typical figures
| Spatial resolution (area mapping) | 10 m (Sentinel-2 NIR/visible); 3 m (Planet SuperDove); 30 m (Landsat 8/9) |
| Spatial resolution (spot verification) | 0.5 m (Planet SkySat, tasked) |
| Revisit cadence | 5 days (Sentinel-2, mid-latitudes); daily (Planet SuperDove); 8 days combined (Landsat 8+9) |
| Typical post-event image latency | 1–7 days cloud-permitting in temperate climates; 1–3 weeks in humid tropical or monsoon seasons |
| Key spectral bands | Red (~665 nm), NIR (~835 nm), red-edge (~705 nm, ~740 nm on Sentinel-2 and SuperDove), SWIR (~1610 nm, ~2190 nm) |
| Minimum resolvable parcel (reliable NDVI) | ~0.5 ha at Sentinel-2 resolution; ~0.05 ha at Planet SuperDove resolution |
| Archive depth | Sentinel-2 from 2015; Landsat from 1982 (USGS); Planet from ~2016 (commercial licence) |
| Primary analytic output | Per-parcel NDVI anomaly score (post-event minus phenology-matched reference) |
| Delivery formats | GeoTIFF raster, GeoJSON or Shapefile polygon layer, ranked CSV of affected policy IDs, PDF loss-adjustment report |
| Cloud cover limitation | Optical sensors only; persistent cloud requires multi-date compositing or SAR supplementation |
Analytics Satellize can run
| Post-event NDVI anomaly map | Pixel-wise NDVI differencing between post-event image and phenology-matched pre-event or multi-year median baseline; radiometric normalisation using pseudo-invariant features | GeoTIFF raster and GeoJSON polygon layer with per-parcel mean anomaly score, delivered within 48 hours of first cloud-free post-event acquisition |
| Damage severity classification | Thresholded NDVI anomaly classified into severity tiers (minor, moderate, severe, destroyed) using crop-type-specific thresholds derived from published validation literature and available ground-truth | Polygon layer attributed with severity class and estimated fractional canopy loss, suitable for direct ingestion into claims management systems |
| Harvest versus damage disambiguation | Dense time-series analysis of NDVI trajectory using all available cloud-free acquisitions in the 30 days before and after the event; abrupt single-date transitions flagged as probable harvest | Per-parcel classification flag (damage probable / harvest probable / ambiguous) with supporting time-series chart for adjuster review |
| Portfolio triage ranked list | Spatial join of damage polygon layer to insured parcel cadastre; parcels ranked by estimated damage severity and area affected | CSV ranked by estimated indemnity exposure, delivered to claims manager within 72 hours of event, enabling prioritised field-visit scheduling |
| Multi-season phenological baseline | Median NDVI compositing across three to five prior growing seasons using Sentinel-2 and Landsat archive; used as normal-year reference for anomaly calculation | Per-crop-type baseline raster stack, updated annually, archived for audit purposes |
| Spot-verification imagery for disputed claims | Tasked SkySat acquisition at 0.5 m resolution over specific disputed parcels; visual interpretation of lodging, scorch, flood scour or physical crop destruction | Annotated high-resolution image report with coordinate-stamped observations, formatted for legal or regulatory submission |
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.