Crop yield nowcasting from multispectral imagery
Multispectral satellites can estimate in-season crop biomass and forecast end-of-season yield weeks before harvest, using vegetation indices tied to chlorophyll and leaf-area physics. This page explains the method, its resolution and cloud-cover limits, and why latency is the critical variable for commodity-price analytics.
Sensors
- Sentinel-2 MSI (ESA): 10 m resolution in visible and NIR bands; 20 m in red-edge (bands 5, 6, 7) and SWIR. Five-day revisit at the equator with both satellites (2A and 2B combined), often three to four days at mid-latitudes. The red-edge bands are the primary reason Sentinel-2 is preferred over Landsat for chlorophyll-sensitive indices such as NDRE.
- Planet SuperDove: 3 m resolution, daily revisit globally. Eight spectral bands including two red-edge channels. High spatial resolution resolves field boundaries below 1 ha, but the constellation requires a commercial licence. Radiometric consistency across the fleet has improved since the SuperDove generation but remains a calibration consideration for multi-year time series.
- Landsat 8/9 OLI (USGS/NASA): 30 m resolution, 16-day revisit per satellite (8-day combined). No dedicated red-edge band, limiting NDRE computation. Strong archive depth back to 1972 (Landsat 1 series) and rigorous cross-calibration make it indispensable for building multi-decade yield baselines. Band 5 (NIR) and Band 4 (red) support standard NDVI.
- MODIS Terra/Aqua (NASA): 250 m resolution for red and NIR bands (bands 1 and 2); 500 m for the remaining spectral bands. Daily global coverage. Useful for national-scale crop monitoring and anomaly detection, but the 250 m floor means individual fields smaller than roughly 50 ha are mixed-pixel problems. The MOD13 vegetation-index product is a well-established operational dataset.
What the canopy is actually broadcasting
Green leaves absorb red light (roughly 620–700 nm) strongly and reflect near-infrared (NIR, roughly 700–1300 nm) strongly. The ratio between those two responses is the basis of NDVI, which has been used operationally since the 1970s. The physics is straightforward: chlorophyll in the mesophyll layer drives the red absorption, and the spongy cell structure of healthy leaves drives the NIR reflectance. Stress, senescence or sparse canopy cover all collapse the ratio.
NDVI saturates at high biomass densities, typically above a leaf area index (LAI) of around 3 to 4. That is a real limit for dense maize or soybean canopies at peak season. The Enhanced Vegetation Index (EVI) was developed partly to address this, incorporating a blue-band correction for atmospheric aerosol scattering and a canopy background adjustment. For crops where saturation is a concern, EVI gives a more linear response across the full growth cycle.
The red-edge region (roughly 700–740 nm) is where the transition from chlorophyll absorption to NIR reflectance happens. Small changes in chlorophyll concentration shift this edge measurably. The Normalised Difference Red Edge index (NDRE), which uses a red-edge band rather than the red band in the numerator, is less prone to saturation and more sensitive to chlorophyll content at high LAI. This is why Sentinel-2's 20 m red-edge bands (B5, B6, B7) are analytically important: they give access to NDRE at a field-relevant scale.
From index to yield: the calibration problem
A vegetation index is a dimensionless number. Converting it to tonnes per hectare requires a calibration layer. The standard approach is to regress in-season index values (typically integrated over the growing season as a cumulative or peak metric) against historical yield statistics from agricultural census data, crop-cut surveys or grain-trade records. The relationship is crop-specific, region-specific, and sensitive to the phenological stage at which the index is measured.
Timing matters enormously. For winter wheat in Northern Europe, NDRE measured at heading (Zadoks scale 55–60) has stronger predictive power than NDVI measured at any single point. For maize, the period around silking (roughly 60–70 days after emergence) is the critical window. Missing that window by even one or two weeks, because of cloud cover or processing latency, degrades the forecast materially.
Calibration models range from simple linear regression to Gaussian process emulators of physically based crop growth models such as DSSAT or APSIM, where satellite-derived LAI is used to update the model state via data assimilation. The physically based route is more transferable across geographies but requires local soil and weather inputs. The empirical regression route is faster to deploy but can fail badly when applied outside its training domain, for example in an anomalous drought year.
Cloud cover is not an excuse, but it is a genuine constraint
Tropical and monsoonal growing regions can sustain cloud cover for weeks at a time, precisely during the crop growth periods that matter most. Sentinel-2's five-day revisit sounds adequate until you account for the fact that a single overcast period in June can eliminate every usable observation during the critical heading window for a winter crop.
Several practical mitigations exist. Temporal compositing, taking the maximum or median index value across a rolling window of observations, reduces cloud gaps but introduces phenological smearing. Gap-filling using MODIS daily observations at coarser resolution can preserve the temporal signal when fine-resolution data is unavailable. Combining optical data with SAR backscatter (Sentinel-1 C-band, which penetrates cloud) can constrain the LAI estimate independently, though SAR-to-LAI relationships are less direct and are covered in the grain-storage reference page rather than here.
Honest summary: for a commodity analyst who needs a field-level yield estimate with a four-week lead time before harvest, cloud cover over a tropical smallholder region is a genuine source of forecast uncertainty that cannot always be resolved. The uncertainty should be quantified and reported, not suppressed.
Resolution floors and the smallholder problem
At 250 m (MODIS), a single pixel covers 6.25 ha. Most smallholder fields in sub-Saharan Africa, South and Southeast Asia are well below 1 ha. A MODIS pixel over such a landscape is a mixture of multiple crops, field boundaries, paths and fallow land. The index it returns is a landscape average, not a crop signal. National-scale anomaly detection is still valid, but field-level yield estimation is not.
Sentinel-2 at 10 to 20 m resolves fields down to roughly 0.5 to 1 ha before mixed-pixel effects become dominant. Planet SuperDove at 3 m resolves most individual smallholder plots. The trade-off is cost and data volume. A full-season Planet tasking programme over a large agricultural region generates substantial data volumes and licence costs. Sentinel-2 is free and open but the 20 m red-edge bands are the limit of what is available without a commercial constellation.
For commodity-price analytics focused on major producing regions (US Corn Belt, Brazilian Cerrado, Ukrainian winter wheat), field sizes are typically large enough that Sentinel-2 is sufficient. For sovereign food-security programmes covering smallholder landscapes, Planet or similar commercial constellations are often necessary.
Latency and the commodity-price use case
A yield forecast published the week before harvest has almost no commodity-market value. The price has already moved. The value of satellite-based nowcasting lies in producing a statistically significant estimate six to ten weeks before harvest, when futures markets are still pricing in uncertainty and before official government crop reports are released.
Latency has two components. Acquisition latency is the time between satellite overpass and delivery of analysis-ready data. For Sentinel-2, ESA's Copernicus Data Space typically delivers Level-2A (atmospherically corrected) products within a few hours of acquisition. Processing latency, running the index calculation, applying the calibration model and generating a field-level output, can be reduced to minutes with a pre-built pipeline. The bottleneck in practice is usually the cloud-gap problem described above, not the processing speed.
Satellize runs this pipeline operationally for the Kingdom of Tonga crop-estimation programme, where small island field sizes and frequent cloud cover make the calibration and gap-filling choices particularly consequential. Analysts who want to discuss the method's applicability to a specific commodity region or sovereign food-security context can request a scope call with the analytics team.
What a yield nowcast can and cannot tell you
A well-calibrated satellite yield model over a major producing region can typically achieve forecast errors in the range of 5 to 15 per cent of mean yield at the administrative-district level, based on published validation studies using Sentinel-2 and Landsat data. At the individual field level, errors are larger. At the national level, aggregation reduces random error but systematic biases from calibration drift or anomalous weather can persist.
The model cannot see below the canopy. A crop that looks green from orbit may be suffering root stress, pest damage or waterlogging that will only appear in the canopy signal weeks later. It cannot distinguish between a high-yield field that is simply at an early growth stage and a low-yield field at peak greenness. Phenological anchoring, knowing where the crop is in its growth cycle, is essential to interpreting the index correctly, and that requires either ground-truth planting-date data or a separate phenology detection algorithm.
Used honestly, a satellite yield nowcast is a probabilistic range estimate, not a point prediction. Presenting it as a single number to a commodity desk or an insurance underwriter without the confidence interval attached is a misuse of the method.
Typical figures
| Best available spatial resolution (field level) | 3 m (Planet SuperDove); 10 m visible / 20 m red-edge (Sentinel-2); 30 m (Landsat 8/9) |
| Minimum field size for reliable signal | ~0.5–1 ha at Sentinel-2 20 m red-edge; ~0.1 ha at Planet 3 m; ~6 ha at MODIS 250 m |
| Revisit frequency | Daily (MODIS, Planet); 3–5 days (Sentinel-2A+2B combined at mid-latitudes); 8 days (Landsat 8+9 combined) |
| Key spectral bands for yield estimation | Red (620–700 nm), Red-edge (700–740 nm), NIR (740–900 nm); blue band for EVI aerosol correction |
| Typical data latency (acquisition to analysis-ready) | 2–6 hours for Sentinel-2 L2A via Copernicus Data Space; same-day for Planet; 1–2 days for Landsat |
| Forecast lead time before harvest | 4–10 weeks for statistically significant district-level estimates in major producing regions |
| Typical district-level forecast error (published literature) | 5–15% of mean yield; higher in smallholder or cloud-affected regions |
| Archive depth | Sentinel-2: from 2015; Landsat: from 1972; MODIS: from 2000; Planet: from ~2016 (SuperDove generation from 2021) |
| Cloud-cover mitigation | Temporal compositing, MODIS gap-fill, multi-sensor fusion; cloud gaps remain a hard constraint in tropical regions |
| Delivery formats | GeoTIFF index rasters, field-polygon CSV/GeoJSON with index statistics, time-series charts, API feed |
Analytics Satellize can run
| In-season NDVI / EVI / NDRE time series per field polygon | Atmospherically corrected surface reflectance (Sentinel-2 L2A or Landsat Collection 2 SR); standard band-ratio index formulas | GeoJSON field layer with per-observation index values and cloud-flag metadata; updated each clear overpass |
| Peak-season LAI estimate | Empirical or physically based inversion of canopy reflectance; Sentinel-2 red-edge bands as primary input | Raster layer (20 m) and polygon summary table; delivered at phenological peak identified by the time-series algorithm |
| District-level yield forecast with confidence interval | Regression or data-assimilation model calibrated against historical yield statistics; cumulative or peak index as predictor | Tabular forecast report (tonnes per hectare, 80% confidence range) per administrative unit; issued at 8-week and 4-week pre-harvest milestones |
| Crop-type mask | Supervised classification using multi-temporal Sentinel-2 index stack and phenological signatures; random forest or similar | Raster crop-type map (10 m) clipped to area of interest; required as input layer before yield modelling |
| Anomaly map relative to 5-year baseline | Z-score of current-season index against same-week historical mean and standard deviation from Sentinel-2 or Landsat archive | Weekly GeoTIFF anomaly layer and summary dashboard; flags districts deviating more than one standard deviation from baseline |
| Phenological stage detection | Time-series curve fitting (e.g. double-logistic or Savitzky-Golay smoothing) to identify green-up, heading and senescence dates | Per-field phenology calendar (CSV/GeoJSON); used to anchor yield model to the correct growth-stage observation window |
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.