Within-field variability mapping for precision agriculture prescriptions
Sub-metre to 3-metre multispectral imagery can resolve within-field nutrient zones that 10-metre sensors miss, but map accuracy and agronomic response are not the same thing. This page explains the physics, the sensors, and the honest limits.
Sensors
- Planet SuperDove: 3-metre resolution, 8 spectral bands including red-edge at 705 nm, daily revisit globally. Radiometric consistency across the constellation is the main calibration challenge for time-series work.
- Maxar WorldView-3: 31-cm panchromatic, 1.24-metre multispectral (8 VNIR bands). Sub-metre detail resolves row-level canopy gaps but revisit is tasking-dependent, typically 1 to 4.5 days at off-nadir, and cloud risk accumulates quickly in humid climates.
- Sentinel-2 MSI: Three red-edge bands (705, 740, 783 nm) at 20-metre resolution, 10-day revisit per satellite (5-day with both Sentinel-2A and 2B). Free and open. At 20 metres, a single pixel covers 400 m²; small fields or narrow management zones are poorly resolved.
- MicaSense Altum (UAV): 5-band multispectral plus thermal, sub-5-cm resolution at typical survey altitude. Not a satellite sensor, but widely used as calibration ground-truth for satellite-derived NDRE and chlorophyll index maps; its thermal channel adds canopy temperature for water-stress cross-checks.
What a canopy reflects, and why nitrogen shows up in the red-edge
Chlorophyll absorbs strongly in the red (around 670 nm) and reflects strongly in the near-infrared (above 750 nm). The slope between those two regions, the red-edge, shifts position and steepness with chlorophyll concentration. Because leaf nitrogen is tightly coupled to chlorophyll content through the photosynthetic apparatus, red-edge reflectance correlates with leaf nitrogen status across a wide range of crop species and growth stages. This is the physical basis for the red-edge normalised difference index (NDRE), calculated as (NIR minus red-edge) divided by (NIR plus red-edge), and for various chlorophyll indices derived from band ratios in the same spectral region.
NDVI, which uses the broad red band rather than the red-edge, saturates at moderate-to-high canopy cover. A dense wheat canopy at flag-leaf stage can return near-identical NDVI values whether it is nitrogen-sufficient or mildly deficient. NDRE saturates later and retains sensitivity into high-biomass conditions, which is precisely when late top-dressing decisions matter most. The practical implication: sensors without a red-edge band are less useful for in-season nitrogen prescription once canopies close.
Resolution floors: what 3 metres sees that 20 metres misses
Agronomically meaningful within-field zones can be as small as a few hundred square metres. A variable-rate spreader working at 12-metre boom width needs input zones at roughly that scale to act on the information. At 20-metre resolution (Sentinel-2 red-edge bands), a single pixel already spans a 20-by-20-metre block; the map can identify broad gradients but not the fine-grained patchiness that drives yield losses in sandy or variable-texture soils.
Planet SuperDove at 3 metres resolves zones down to roughly 9 m², though the effective minimum mappable unit for reliable statistics is closer to 10 to 15 pixels, so about 90 to 135 m². WorldView-3 at 1.24-metre multispectral resolution pushes that floor lower still, but its eight VNIR bands do not include a dedicated red-edge channel; the 705 nm red-edge is absent from the standard WorldView-3 band set, which limits its direct NDRE utility compared with SuperDove or Sentinel-2.
UAV-mounted sensors like the MicaSense Altum produce sub-5-cm imagery that resolves individual plant rows. That resolution is rarely needed for prescription maps, but it is indispensable for validating whether satellite-derived indices are tracking the right signal or picking up soil background, shadow, or mixed-pixel artefacts at field edges.
Temporal frequency: one image is a snapshot, not a decision
Nitrogen uptake in cereals is not linear. There are two or three critical windows per season when top-dressing alters final grain protein and yield, and those windows can be as narrow as seven to ten days. A single clear-sky acquisition may land outside the agronomically relevant period entirely. Cloud cover compounds this: in temperate maritime climates, the probability of obtaining a usable optical image within any given seven-day window can fall below 40 percent in spring.
Planet's daily revisit makes it the most likely constellation to deliver a usable acquisition within a narrow agronomic window, but daily revisit does not mean daily clear-sky imagery. Sentinel-2's five-day combined revisit, combined with its free archive back to 2015, supports multi-year baseline construction and anomaly detection. The practical workflow for prescription mapping typically fuses both: Sentinel-2 for temporal context and trend, SuperDove for the high-resolution snapshot closest to the application date.
Latency matters too. A prescription map delivered 72 hours after image acquisition is still useful for a planned spreading operation. A map delivered ten days later is not. Processing pipelines need to run close to real time, with atmospheric correction, cloud masking, and index computation completed within 24 hours of image downlink.
The gap between a good map and a good agronomic outcome
This is where buyers often receive an incomplete picture. A spatially accurate NDRE map tells you where the canopy is nitrogen-stressed relative to the rest of the field. It does not tell you why. Stress can reflect soil texture variation, waterlogging, compaction, pest damage, or historical management differences, all of which produce similar spectral signatures. Applying additional nitrogen to a waterlogged patch will not fix the waterlogging.
The conversion from index value to a kilogram-per-hectare prescription also introduces uncertainty. Published algorithms for this conversion, such as those embedded in the Yara N-Tester or Hydro N-sensor approaches, were developed under specific crop varieties, climates, and baseline fertility conditions. Transferring them to a different context without local calibration data can produce prescriptions that are directionally correct but quantitatively unreliable. The honest position is that satellite-derived prescription maps reduce input waste and improve within-field uniformity, but the magnitude of yield or protein response depends heavily on agronomic context that the satellite cannot observe.
Ground-truth sampling, even at modest density (one composite soil or tissue sample per management zone), substantially improves prescription confidence. Satellite data identifies the zones; soil and tissue analysis validates what is driving them.
Building the prescription layer: from index to variable-rate file
The standard workflow starts with atmospherically corrected surface reflectance. For Sentinel-2, the Sen2Cor processor produces Level-2A products; for Planet, TOAR (top-of-atmosphere reflectance) is the standard delivery, with surface reflectance requiring additional correction. NDRE and chlorophyll index layers are computed per pixel, then segmented into management zones using unsupervised clustering (k-means is common) or fuzzy-c-means to handle the gradual transitions typical of soil-driven variability.
Zone boundaries are simplified to polygons compatible with variable-rate technology controllers, which typically accept ISOXML or shapefile formats. Each zone is assigned a rate modifier relative to a farm-specific baseline rate, not an absolute kilogram figure, because the satellite cannot determine the baseline independently. The prescription file is then loaded to the spreader or sprayer controller before the field operation.
Satellize applies this workflow operationally on open and commercial imagery; its crop-estimation programme for the Kingdom of Tonga demonstrated that index-to-outcome pipelines can be calibrated and delivered in low-infrastructure environments. The same pipeline architecture underpins within-field prescription work, with zone segmentation and rate-modifier outputs delivered as GIS layers ready for farm management software ingestion.
What the method cannot do
Optical sensors of any resolution cannot see through cloud. In regions or seasons with persistent cloud cover, synthetic aperture radar can track broad biomass proxies but lacks the spectral specificity for nitrogen-status mapping. Hyperspectral sensors, which resolve the red-edge in tens of narrow bands rather than two or three, provide better nitrogen quantification but are currently limited to airborne or small-swath satellite platforms; PRISMA (ASI) and DESIS (DLR) exist but have narrow swaths and limited revisit. Spaceborne hyperspectral at field-relevant resolution and revisit remains a near-term capability gap.
Sub-field soil moisture variation, which interacts with nutrient availability, is not directly observable in multispectral reflectance. Canopy temperature from thermal sensors (Landsat 8/9 TIRS at 100-metre resolution, or UAV thermal) adds a water-stress dimension but at coarser spatial resolution than the multispectral prescription map. Integrating both is methodologically sound; it requires careful co-registration and resampling.
Typical figures
| Spatial resolution (operational) | 0.31 m pan / 1.24 m MS (WorldView-3); 3 m (Planet SuperDove); 10–20 m (Sentinel-2 MSI) |
| Revisit frequency | Daily (Planet); 5-day combined (Sentinel-2A+2B); tasking-dependent (WorldView-3, typically 1–4.5 days off-nadir) |
| Key spectral bands | Red-edge: 705 nm (SuperDove, Sentinel-2 B5), 740 nm (Sentinel-2 B6), 783 nm (Sentinel-2 B7); NIR 842 nm (Sentinel-2 B8) |
| Minimum mappable zone (reliable statistics) | ~90–135 m² at 3 m resolution (10–15 pixels); ~4,000 m² at 20 m resolution |
| Processing latency (index to delivery) | Target <24 hours from image acquisition for time-critical applications |
| Archive depth | Sentinel-2 from 2015; Landsat from 1972 (USGS archive); Planet from ~2016 |
| Prescription file formats | ISOXML (ISO 11783), shapefile, GeoTIFF zone raster |
| Cloud cover constraint | Optical sensors fully blocked by cloud; usable-image probability in temperate spring can fall below 40% per 7-day window |
| Radiometric calibration requirement | Surface reflectance (Level-2) mandatory; TOAR alone introduces cross-date index inconsistency |
Analytics Satellize can run
| NDRE and chlorophyll index rasters | Band-ratio computation on atmospherically corrected surface reflectance (Sen2Cor for Sentinel-2; empirical line or 6S for Planet) | GeoTIFF per acquisition date, cloud-masked, co-registered to field boundary |
| Within-field management zone map | Unsupervised clustering (k-means or fuzzy-c-means) on single-date or multi-date index stack | Polygon shapefile with zone IDs and mean index values per zone |
| Variable-rate fertiliser prescription file | Zone-to-rate-modifier lookup calibrated against client baseline rate; ISOXML encoding for VRT controller | ISOXML prescription file ready for field controller upload |
| In-season index trend report | Time-series analysis of NDRE across growing season using Sentinel-2 and Planet fusion | PDF or dashboard showing per-zone index trajectory against seasonal baseline |
| Anomaly detection alert | Statistical threshold on deviation from multi-year median index per pixel (z-score or percentile method) | Geofenced alert with anomaly polygon and probable cause classification |
| UAV calibration cross-check report | Regression of MicaSense Altum sub-5-cm NDRE against satellite-derived NDRE at coincident date | Calibration coefficient table and residual map identifying mixed-pixel or soil-background artefacts |
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.