Winter cover crop species discrimination using phenological timing differences
Common winter cover crops look nearly identical in any one Sentinel-2 image. Their phenological timing differs enough to discriminate species reliably, which matters for carbon credit verification and nutrient auditing.
Sensors
- Sentinel-2 MSI (ESA Copernicus): 10 m resolution in visible and near-infrared bands, 20 m in red-edge and SWIR. Nominal 5-day revisit at the equator with both satellites; in practice 10 days per satellite, often reduced to 3-5 days at mid-latitudes with both. The red-edge bands (B5, B6, B7 at 705, 740, 783 nm) are the primary discriminators for canopy greenness and chlorophyll concentration across the winter season.
- Planet SuperDove: 3 m resolution, daily revisit in most mid-latitude regions, 8 spectral bands including two red-edge channels. High revisit rate is valuable for filling cloud gaps in the Sentinel-2 time series and for resolving abrupt phenological transitions such as radish collapse after a hard frost. Commercially licensed; not freely available.
- Landsat 8/9 OLI (USGS/NASA): 30 m resolution, 16-day revisit per satellite (8-day combined). Lacks dedicated red-edge bands, which limits chlorophyll sensitivity relative to Sentinel-2, but the long archive back to 1984 (Landsat 5 TM) and consistent radiometric calibration make it useful for multi-year phenological baseline construction.
- Copernicus HR-VPP (High-Resolution Vegetation Phenology and Productivity): Derived Sentinel-2 product suite at 10 m, providing pre-computed phenological metrics: start of season, end of season, peak NDVI, and large-scale integral. Updated annually. Useful as a ready-made phenological fingerprint layer but computed over full growing seasons; winter cover crop windows require custom sub-season extraction.
Why a single image cannot tell rye from vetch
In November, a well-established cereal rye field and a vetch field can share NDVI values within 0.05 of each other. Both are dense, green and actively photosynthesising. Their reflectance in the red-edge bands overlaps substantially. Single-date classification trained on spectral signatures alone routinely confuses them, and adding more spectral bands does not resolve the ambiguity because the physiological states are genuinely similar at that moment.
The discriminating information is not in the spectrum at any given date. It is in the shape of the spectral trajectory across the whole winter. Each species has a characteristic phenological fingerprint driven by its biology, and that fingerprint is far more stable than its instantaneous reflectance.
What each species does to the spectral record across winter
Radish (Raphanus sativus) is the most tractable species from a remote-sensing standpoint precisely because it is so dramatic. It is not winter-hardy in temperate climates with sustained sub-zero temperatures. After a hard frost, typically in December or January depending on latitude and year, the canopy collapses within days. NDVI drops from roughly 0.6-0.7 to below 0.2, and the field becomes spectrally indistinguishable from bare soil. This abrupt disappearance is a reliable phenological marker. No other common cover crop species does this.
Cereal rye (Secale cereale) is the dominant winter-hardy species. It greens up early in autumn, maintains moderate NDVI through the coldest months (often 0.4-0.6 even in January at 50°N), then accelerates growth rapidly in late February and March as temperatures rise. Senescence and termination by the farmer typically produce a sharp NDVI decline in April or May before cash-crop planting.
Legumes, including hairy vetch (Vicia villosa) and crimson or white clover (Trifolium spp.), green up more slowly in autumn than rye and maintain greenness later into spring. Vetch in particular continues vigorous growth well into May if left unterminated, and its red-edge reflectance remains elevated longer than rye. The later end-of-season date is a consistent discriminator.
Clover species are the most spectrally similar to each other and to vetch during peak greenness. Discrimination within the legume group requires either very high revisit frequency to resolve subtle timing differences, or ancillary data such as field management records. Honest assessment: satellite phenology alone will often collapse clover species into a single legume class rather than resolving individual species.
Building a usable time series through cloud gaps
The Sentinel-2 10-day revisit sounds adequate until you account for cloud cover. In northern Europe and the US Corn Belt, winter cloud frequency can leave stretches of three to five weeks with no usable observations. A phenological classifier that relies on a smooth, densely sampled curve will fail if the curve has large gaps at critical transition points, such as the frost-kill window for radish.
The standard mitigation approaches are well documented in the literature. Harmonic regression (fitting sinusoidal functions to the available observations) smooths over gaps but can distort abrupt transitions. Savitzky-Golay filtering is less prone to distortion at step changes. Gap-filling using Planet SuperDove or Landsat observations as ancillary inputs improves temporal density but introduces inter-sensor calibration uncertainty that must be corrected. The Copernicus HR-VPP products apply their own smoothing pipeline, which is appropriate for full-season crops but may miss the short radish senescence window.
In practice, a minimum of 8-12 cloud-free observations across the October-to-May window is needed for reliable species-level classification. Years with persistent winter cloud cover over a given region will produce lower classification confidence, and any honest delivery of results should include a per-field data-quality flag indicating observation density.
Turning phenological curves into species labels
The classification approach that consistently outperforms single-date methods in published work is time-series feature extraction followed by supervised machine learning. The features extracted from the smoothed NDVI or red-edge index curve include: date of green-up onset, peak index value and its date, rate of autumn green-up, rate of spring green-up, date of senescence onset, and end-of-season date. These six to eight scalar features per field replace the raw time series as inputs to a random forest or gradient-boosted classifier.
Training data requirements are modest by machine-learning standards. A few hundred ground-truth field observations per species, collected across two or three seasons to capture inter-annual climate variability, is sufficient to train a classifier that generalises across a region. The classifier's confidence is highest for radish (the frost-kill signal is unambiguous) and for rye versus the legume group. Within-legume discrimination remains the hardest problem.
Sentinel-2's red-edge bands add meaningful separability beyond NDVI alone. The Chlorophyll Red-Edge index (CIre, using B7 and B5) tracks canopy chlorophyll concentration more sensitively than broadband NDVI, which is particularly useful for distinguishing the slower autumn green-up of legumes from rye during October and November when NDVI differences are small.
Agronomic applications: carbon credits and nutrient accounting
Cover crop species identity is not a curiosity. It determines the nitrogen contribution to the following cash crop, the carbon sequestration rate claimable under voluntary carbon market protocols, and the erosion protection provided over winter. Rye and legumes differ substantially in all three. A vetch or clover field can fix 80-200 kg of atmospheric nitrogen per hectare per season under good conditions; rye fixes none. Carbon credit registries increasingly require evidence of species and termination date, not just evidence that something green was planted.
Satellite-derived species maps provide a scalable verification layer that farm visits cannot. A single agronomist can inspect perhaps 20 fields per day. A Sentinel-2 classification covering 50,000 fields costs the same computational effort regardless of field count. The honest caveat is that satellite data verifies what was growing, not why it was growing or whether it was terminated correctly. Integration with field-level management records and soil sampling remains necessary for full protocol compliance.
Satellize has applied phenological time-series methods to crop classification in its Tonga programme, and the same pipeline is adaptable to cover crop species work in temperate agricultural regions. For clients building carbon credit verification workflows or national nutrient management audit systems, the relevant next step is a feasibility assessment against your specific geography, target species mix, and cloud climatology.
Honest limits of the method
Small fields are a genuine problem. Sentinel-2 at 10 m means that a field narrower than roughly 30-40 m will have its edge pixels contaminated by adjacent bare soil or hedgerow, depressing NDVI and distorting the phenological curve. Planet SuperDove at 3 m resolves this but at commercial cost.
Mixed-species cover crop blends, which are agronomically common, produce intermediate spectral curves that the classifier may assign to the dominant species or to no species at all. The method works best on single-species or two-species plantings where one component is clearly dominant by biomass.
Inter-annual climate variability shifts phenological dates by two to four weeks. A classifier trained on one or two seasons may underperform in an anomalously warm or cold year. Retraining or recalibrating on recent observations each season is advisable, not a one-time deployment.
Typical figures
| Primary sensor spatial resolution | 10 m (Sentinel-2 visible/NIR), 20 m (red-edge/SWIR) |
| Revisit frequency (Sentinel-2, both satellites) | 3-5 days at mid-latitudes; 10 days per individual satellite |
| Minimum field size for reliable classification | Approximately 0.5 ha at 10 m resolution; smaller fields carry edge-pixel contamination risk |
| Spectral bands used | Red-edge (705, 740, 783 nm), NIR (842 nm), SWIR (1610 nm); NDVI and CIre indices derived |
| Minimum usable observations per season | 8-12 cloud-free acquisitions across October-May window |
| Classification accuracy (rye vs. legume group) | Typically 80-90% overall accuracy in published studies with adequate training data; within-legume species separation lower |
| Archive depth (Sentinel-2) | From 2015 (Sentinel-2A launch); Landsat extends usable archive to 1984 |
| Phenological metric latency | End-of-season metrics available 2-4 weeks after season close; near-real-time frost-kill detection within 10-15 days of event |
| Delivery formats | GeoTIFF species classification raster, per-field species probability table (GeoPackage or CSV), confidence/data-quality flags |
Analytics Satellize can run
| Per-field winter cover crop species map | Phenological feature extraction (green-up date, peak NDVI, senescence date) from smoothed Sentinel-2 NDVI and CIre time series, classified by random forest | GeoTIFF raster and field-polygon attribute table with species label and per-class probability score |
| Radish frost-kill detection alert | Threshold-based NDVI drop detection on sequential Sentinel-2 acquisitions following sub-zero temperature events (cross-referenced with ERA5 reanalysis temperature fields) | Field-level alert layer with estimated frost-kill date and confidence flag, delivered within 10-15 days of event |
| Seasonal phenological curve archive per field | Savitzky-Golay smoothed NDVI and red-edge index time series extracted per field polygon from full Sentinel-2 archive | CSV time series per field, suitable for input to carbon credit protocol documentation or agronomic audit records |
| Cloud-gap-filled composite time series | Harmonic regression gap-filling on Sentinel-2 observations, with optional Planet SuperDove or Landsat 8/9 ancillary integration and inter-sensor cross-calibration | Regularised 10-day interval NDVI raster stack across the October-May window |
| Species-level nitrogen credit estimate | Lookup of published species-specific nitrogen fixation ranges (legumes) and biomass accumulation curves (rye) applied to satellite-derived species map and green-area index proxy | Per-field nitrogen contribution table with uncertainty range, formatted for nutrient management audit submission |
| Multi-year phenological baseline and anomaly flag | Year-on-year comparison of extracted phenological metrics against a 5-year Sentinel-2 baseline; anomaly flagging for fields whose green-up or senescence dates fall outside one standard deviation of the baseline | Annual anomaly report identifying fields where declared species is inconsistent with observed phenological behaviour |
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.