Harvest progress monitoring at regional scale

Harvest is the fastest phenological transition in the agricultural calendar, and missing it costs governments and traders dearly. Sequential optical and SAR observations detect the abrupt canopy-removal signal, distinguish harvest from crop failure, and map regional progress at sub-weekly cadence.

Sensors

Planet SuperDove: 3 m ground sampling distance, 8 spectral bands including red-edge, daily revisit at mid-latitudes from a constellation of roughly 200 satellites. The combination of daily cadence and sub-5 m resolution makes it the only open-commercial option capable of resolving individual field harvest events within a 24-hour window, though archive access requires a commercial licence.
Sentinel-2 MSI: 10 m resolution in visible and near-infrared bands, 20 m in red-edge and shortwave infrared, 5-day revisit at the equator with both satellites (2A and 2B). SWIR bands 11 and 12 are particularly sensitive to residue moisture and soil exposure after cutting. Cloud cover is the primary operational constraint.
Sentinel-1 SAR (C-band): IW mode delivers 10 m range resolution, 6-day single-satellite revisit (3 days with both 1A and 1C). C-band backscatter drops markedly when the volume-scattering wheat or maize canopy is removed and flat, dry stubble replaces it. SAR operates through cloud, which is its decisive advantage during monsoon or maritime harvest seasons.
MODIS Terra/Aqua: 250 m resolution NDVI in bands 1 and 2, daily or near-daily revisit. Too coarse to resolve individual fields in fragmented landscapes, but useful for regional-scale harvest progress curves across large homogeneous breadbaskets such as the North American Great Plains or the Ukrainian steppe. 20-plus years of archive enables robust climatological baselines.

What the signal actually looks like

Harvest is, spectrally speaking, violent. A standing wheat canopy at grain-fill carries NDVI values typically between 0.6 and 0.8. Within 24 to 48 hours of a combine passing through, the same pixel returns NDVI values of 0.1 to 0.3, depending on residue cover and soil moisture. The Normalised Difference Residue Index (NDRI), which uses SWIR and near-infrared, responds even more sharply because freshly cut straw has a distinctive shortwave-infrared reflectance signature quite different from either live vegetation or bare soil.

SAR tells a parallel story. A dense cereal canopy in C-band exhibits volume scattering and relatively high cross-polarised (VH) backscatter. After harvest the surface reverts to a rough, low-moisture stubble or bare soil, and VH backscatter typically drops by 3 to 6 dB, a change large enough to detect reliably against background noise. The VV channel is more sensitive to surface roughness and soil moisture, so the ratio VH/VV shifts characteristically. This is not a subtle signal. The challenge is not detection but interpretation.

Harvest or failure: the ambiguity that matters most

A field that goes from green to brown in one image interval looks identical whether a combine harvested it or a drought killed it. This is the central interpretive problem, and resolving it requires context, not just imagery.

Thermal accumulation models constrain the answer. Growing degree days (GDD), calculated from daily maximum and minimum air temperatures against a crop-specific base temperature (10°C for maize, 0°C for winter wheat in most published formulations), define the physiologically plausible harvest window. A canopy-removal event occurring within the expected GDD range for the crop type in that pixel is almost certainly harvest. The same signal occurring 200 to 400 GDD early, during a documented heat or drought episode, is almost certainly stress-induced senescence or crop failure. ERA5 reanalysis temperature data, available from the Copernicus Climate Data Store, provides the thermal accumulation layer at 0.25-degree resolution globally.

Pre-harvest NDVI trajectory adds further discrimination. A healthy crop approaching harvest shows a characteristic green-up followed by gradual senescence before the abrupt removal event. A stressed crop shows an earlier, slower decline with no sharp terminal drop. Distinguishing these trajectories requires dense time series, which is exactly where Planet's daily cadence or Sentinel-1's cloud-penetrating revisit earns its cost.

Revisit arithmetic and why it matters for fast campaigns

Wheat harvest in a temperate breadbasket can progress at 5 to 15 percent of total area per day during peak campaign. A regional monitoring system with a 10-day revisit will see only the start and end states, missing the spatial pattern of progress entirely. At 5-day revisit (Sentinel-2 single-pass) you catch the broad shape of the campaign but misdate individual field events by up to four days. At daily revisit (Planet or MODIS) you can map the harvest frontier as it moves.

The practical answer for most government clients is sensor fusion. Sentinel-1 provides the cloud-immune backbone at 3-day revisit, anchoring the temporal structure. Sentinel-2 adds spectral richness when skies clear, allowing residue type and soil condition to be characterised. Planet fills the gaps in fragmented landscapes where 10 m is too coarse to separate small fields. MODIS provides the regional summary curve that communicates progress to non-specialist audiences. No single sensor does all of this.

Turning field events into regional statistics

The operational output is a harvest-progress curve: percentage of planted area harvested by date, stratified by crop type and administrative unit. Producing it requires three upstream steps. First, a crop-type mask from the season's classification must exist; without knowing which pixels are wheat versus sunflower versus fallow, a canopy-removal signal is uninterpretable. Second, change detection must be applied to the time series, typically using a threshold on NDVI drop rate or a SAR backscatter change exceeding a calibrated minimum. Third, each detected event must be assigned a harvest date, usually the midpoint between the last pre-harvest and first post-harvest observation.

Uncertainty is real and should be reported. Date uncertainty equals half the revisit interval. Area uncertainty depends on the accuracy of the upstream crop-type mask, which in fragmented landscapes can carry 10 to 20 percent commission errors for minority crop types. Any honest regional progress report should carry these confidence intervals rather than presenting a single number as fact.

Operational constraints worth knowing before you commission

Cloud cover is the most common cause of missed harvest events in optical time series. Persistent cloud during harvest season, common in South and South-East Asia during the kharif season, can produce data gaps of two to three weeks in Sentinel-2 records. SAR is the mitigation, not a complete solution: SAR cannot distinguish crop type without an optical anchor, and wet stubble after rain can temporarily restore backscatter values that mimic a standing canopy.

Spatial resolution sets a minimum field size below which detection becomes unreliable. At 10 m resolution (Sentinel-2), fields smaller than roughly 0.5 hectares begin to suffer from mixed-pixel contamination at field edges. At 250 m (MODIS), the minimum interpretable unit is around 25 hectares, which excludes most smallholder landscapes in sub-Saharan Africa or South-East Asia entirely. This is not a solvable problem with processing alone; it requires a higher-resolution sensor.

Satellize runs this kind of dense time-series analysis on open constellations for government clients, including crop-estimation work for the Kingdom of Tonga, and can add commercial Planet tasking where the revisit or resolution requirements exceed what free-tier data supports.

What the archive makes possible beyond the current season

Sentinel-2 archive runs to 2015, Landsat to 1972, and MODIS to 2000. This depth allows harvest progress curves to be placed in climatological context: is this year's campaign running two weeks early, and is that unusual? Over a 20-year MODIS record, the interannual variability of harvest timing in a given region can be characterised, and anomalies can be flagged with statistical confidence rather than anecdote.

For food security applications, the most useful product is often not the current-year map but the comparison: how does this season's pace compare to the five-year average, and what does historical evidence suggest about the relationship between early harvest and final yield? That question sits at the boundary between harvest monitoring and yield forecasting, which is addressed separately in the crop yield estimation page.

Typical figures

Best spatial resolution (optical)	3 m (Planet SuperDove); 10 m (Sentinel-2 visible/NIR)
Best spatial resolution (SAR)	10 m (Sentinel-1 IW mode)
Minimum detectable field size	~0.1 ha at 3 m; ~0.5 ha at 10 m; ~25 ha at 250 m (MODIS)
Revisit cadence	Daily (Planet, MODIS); 3 days (Sentinel-1 dual-satellite); 5 days (Sentinel-2 dual-satellite)
Key spectral bands	NIR, red-edge, SWIR1/2 (optical); C-band VV and VH (SAR)
Harvest date uncertainty	±0.5 × revisit interval (typically ±1 to ±2.5 days at best cadence)
Latency from acquisition to analysis	6–24 hours for Sentinel open data; 12–48 hours typical for commercial pipeline
Archive depth	MODIS: 2000–present; Sentinel-2: 2015–present; Landsat: 1972–present
Coverage	Global; polar regions limited by sun angle in optical sensors
Delivery formats	GeoTIFF change maps, vector field polygons with harvest date attribute, CSV progress curves, PDF bulletins

Analytics Satellize can run

Harvest-progress curve by crop type and admin unit	Threshold-based change detection on NDVI and SWIR time series, stratified by crop-type mask	Weekly CSV and chart showing percent area harvested by district and crop
Per-field harvest date map	Breakpoint detection (e.g. BFAST or CCDC framework) applied to dense optical or SAR time series	Vector polygon layer with ISO-date attribute for each detected harvest event
Harvest vs. crop-failure discrimination layer	GDD-constrained event classification combining ERA5 thermal accumulation and pre-harvest NDVI trajectory shape	Raster map with per-field classification: harvested / stress-senesced / ambiguous, with confidence score
SAR-based cloud-immune harvest detection	Sentinel-1 VH backscatter change detection; threshold calibrated against known harvest events in training region	Binary change raster updated at each 3-day overpass, merged into optical product where cloud-free data exists
Anomaly report vs. climatological baseline	Current-season progress curve compared to 10-year MODIS or Sentinel-2 archive percentile envelope	PDF bulletin flagging regions running more than one standard deviation early or late
Residue cover classification post-harvest	NDRI and SWIR-based index applied to first post-harvest clear observation	Raster layer classifying fields as high / moderate / low residue cover, informing soil carbon and tillage analysis

Sources

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.