Sensors
- Sentinel-2 MSI: 10 m resolution in visible and near-infrared bands (B02, B03, B04, B08); 20 m in red-edge and SWIR (B05, B06, B07, B8A, B11, B12). Five-day revisit at mid-latitudes with two satellites. Free archive from 2015. The red-edge bands (705 nm, 740 nm, 783 nm) are particularly sensitive to the differential crop vigour that expresses buried ditches and pits.
- PlanetScope SuperDove: 3 m resolution, eight bands including two red-edge channels. Near-daily revisit globally. Commercial archive from approximately 2016. Daily cadence is the critical advantage: it captures the narrow phenological window, often only a few days wide, when crop stress over buried features peaks before harvest removes the signal entirely.
- WorldView-3 VNIR/SWIR: 0.31 m panchromatic, 1.24 m multispectral VNIR (eight bands), 3.7 m SWIR (eight bands). Used here for calibration rather than wall-to-wall mapping. Sub-metre resolution allows individual crop-mark widths to be measured and used to constrain the spatial resolution requirements of the broader Sentinel-2 and PlanetScope analysis.
- Landsat 8/9 OLI: 30 m resolution, eight reflective bands including SWIR-1 and SWIR-2. Sixteen-day revisit per satellite, eight days combined. Free archive from 1972 (earlier missions). Useful for extending the temporal baseline before the Sentinel-2 and PlanetScope epochs and for cross-calibrating the long-term spectral record.
Why a single season's imagery has no business in a planning report
Crop-mark expression depends on a chain of contingencies: the crop species, the soil moisture deficit at the critical growth stage, the depth and fill of the underlying feature, and the precise phenological moment when stress over a ditch or pit diverges visibly from stress over undisturbed subsoil. Miss any link and the feature is invisible. A wet spring can suppress the entire signal. A crop harvested early removes it. This is not a theoretical concern. English Heritage's systematic aerial survey programme accumulated decades of observations precisely because no single flight season was sufficient to characterise a landscape.
Multitemporal stacking addresses this directly. By compositing spectral indices across many acquisition dates and multiple growing seasons, each pixel accumulates evidence. A feature that appears faintly in 2017, clearly in 2020, and again in 2022 scores high on the probability surface regardless of what 2018 and 2019 showed. The result is a density map that reflects the statistical weight of the archive rather than the accident of last summer's weather.
Building the probability surface: method in plain terms
The workflow begins with radiometric normalisation across all acquisitions. Surface reflectance products (Sentinel-2 Level-2A, PlanetScope SR) are used throughout. For each date, a crop-stress index is computed. The Normalised Difference Vegetation Index (NDVI) is the standard starting point, but red-edge indices such as the Sentinel-2 Red-Edge Chlorophyll Index (S2RECl) and the Canopy Chlorophyll Content index are more sensitive to the subtle vigour differences that shallow features produce. Soil-mark seasons, when bare earth is tilled, use different indices: the Clay Minerals Ratio (SWIR-1/SWIR-2) and the Iron Oxide Ratio (Red/Blue) are published indicators of disturbed subsoil composition.
Each seasonal composite is then binarised against a locally adaptive threshold, producing a presence-absence layer for that season. The layers are summed across the archive. Pixels that register positive across multiple independent seasons receive higher scores. A Gaussian smoothing kernel is applied to account for the positional uncertainty inherent in 10 m pixels relative to features that may be only 2 to 3 m wide. The output is a continuous probability raster, typically delivered at 10 m, with a categorical overlay classifying cells as low, medium, high or very high feature density. Zones scoring very high across three or more independent seasons are flagged for targeted ground investigation.
What the sensors can and cannot resolve
Sentinel-2 at 10 m will not resolve individual features narrower than roughly 15 to 20 m with confidence. A Roman ditch 2 m wide is below the detection floor and will only appear in the probability surface if it is part of a denser complex whose aggregate spectral anomaly is large enough to shift a 10 m pixel. PlanetScope at 3 m improves this materially: features of 5 to 8 m width become detectable under good conditions. WorldView-3 at 1.24 m multispectral can resolve individual ditches of 3 to 4 m width, which is why it is used for calibration rather than coverage.
Cloud is an unavoidable constraint in temperate climates. A single growing season in northern Europe may yield fewer than ten usable cloud-free acquisitions over a given field. This is precisely why a multi-year archive matters: the effective sample size grows with time, not with any single year's luck. SWIR bands (Sentinel-2 B11, B12; WorldView-3 SWIR) partially penetrate thin cirrus and haze, extending the usable acquisition count. Deep clay soils suppress crop-mark expression even in dry years because capillary moisture persists longer. Sandy and gravelly soils over chalk or limestone are the most responsive substrates, and this geological context should be stated explicitly in any developer impact report.
Calibrating the map against ground truth
A probability surface without calibration is an unvalidated index. Calibration requires comparison against known features: geophysical survey results (magnetometry, earth resistance, ground-penetrating radar), excavation records from earlier phases, and where available, the Historic Environment Record. WorldView-3 tasking over sample areas provides the spatial resolution needed to match individual anomalies in the probability surface to known feature locations and measure false-positive and false-negative rates.
In practice, a well-calibrated surface from a five-to-seven year Sentinel-2 archive, combined with PlanetScope daily data for the most recent two to three seasons, achieves detection rates for moderate-to-large feature complexes that are broadly consistent with published comparisons between satellite and aerial photographic prospection. The honest caveat is that isolated small features, single post-holes, and narrow slot trenches remain below the detection floor of any current multispectral satellite system. The probability surface is a density map, not a substitute for targeted geophysical survey. It tells a developer where to concentrate that survey, and where the risk of encountering significant archaeology is low enough to inform a proportionate mitigation strategy.
What the developer's planning team actually receives
The primary deliverable is a georeferenced raster and vector GIS package: the continuous probability surface, the categorical density classification, and a polygon layer of high-confidence anomaly clusters with per-polygon statistics summarising the number of positive seasons, the peak index value, and the estimated minimum feature width. A written technical note explains the method, the archive dates used, the cloud-cover statistics per season, and the substrate geology context.
Satellize has applied related multitemporal crop-stress compositing methods in its Tonga crop-estimation programme, where stacking across phenological stages similarly resolves signal that no single date captures. The underlying compositing and anomaly-extraction pipeline transfers directly to archaeological feature-density work. Outputs are delivered as GeoTIFF rasters, GeoPackage vectors, and a PDF technical report formatted to meet the evidential standards expected by local planning authorities and their archaeological advisers. Recommended next steps for targeted geophysical survey are mapped spatially, not listed generically.
Archive depth and the long-term planning value
The Sentinel-2 archive runs from June 2015. PlanetScope's usable archive for most of Europe and North America extends to 2016 or 2017 at useful cadence. Together, that is roughly eight to nine growing seasons at the time of writing. Each additional season adds independent evidence. A development project with a long planning timeline benefits from a living analysis: the probability surface is updated annually as new growing seasons complete, and zones that cross a significance threshold trigger alerts.
This longitudinal value is rarely discussed in heritage impact assessments, which typically commission a single study and treat it as definitive. Ploughzone archaeology is dynamic. Repeated cultivation progressively truncates buried features, changing their spectral expression over time. A surface computed in 2025 will differ from one computed in 2030, and that difference is itself informative about the rate of attrition. For major infrastructure projects with phased construction programmes spanning a decade, the archive-update model is not optional; it is the only honest way to track what is being lost.
Typical figures
| Spatial resolution (mapping layer) | 10 m (Sentinel-2 probability surface); 3 m (PlanetScope anomaly overlay) |
| Spatial resolution (calibration) | 1.24 m multispectral, 0.31 m panchromatic (WorldView-3) |
| Revisit cadence | 5 days (Sentinel-2, two satellites); near-daily (PlanetScope SuperDove) |
| Archive depth | Sentinel-2 from June 2015; PlanetScope usable from approx. 2016; Landsat from 1972 |
| Spectral bands used | Visible (Blue, Green, Red), Red-edge (705, 740, 783 nm), NIR (842 nm), SWIR-1 (1610 nm), SWIR-2 (2190 nm) |
| Minimum detectable feature width (satellite) | Approx. 5–8 m under good conditions (PlanetScope); 15–20 m (Sentinel-2); 3–4 m (WorldView-3 VNIR) |
| Cloud-cover constraint | Typically 5–15 usable acquisitions per growing season in temperate NW Europe; multi-year stacking compensates |
| Substrate sensitivity | Highest on sandy/gravelly soils over chalk or limestone; suppressed on deep clays |
| Delivery formats | GeoTIFF (probability raster), GeoPackage (anomaly polygons), PDF technical report |
| Latency (initial study) | 6–10 weeks from data order to delivered GIS package, depending on archive processing volume |
Analytics Satellize can run
| Multitemporal crop-stress probability surface | Per-pixel summation of binarised red-edge and NDVI anomaly layers across all cloud-free acquisitions in the archive; locally adaptive thresholding | GeoTIFF raster (continuous 0–1 probability), updated annually |
| Categorical feature-density classification | Quantile or natural-breaks classification of the probability surface into low/medium/high/very-high zones, with minimum-mapping-unit filtering | GeoPackage polygon layer with per-zone statistics; PDF map for planning report appendix |
| High-confidence anomaly cluster polygons | Connected-component labelling on cells exceeding threshold in three or more independent seasons; per-polygon summary statistics (peak score, season count, estimated width) | GeoPackage vector layer; recommended geophysical survey prioritisation map |
| Soil-mark composite (bare-earth seasons) | Clay Minerals Ratio and Iron Oxide Ratio composites from SWIR and visible bands during tillage windows; anomaly extraction against local soil background | Supplementary GeoTIFF and anomaly overlay integrated with crop-mark surface |
| WorldView-3 calibration report | Sub-metre multispectral feature delineation over sample areas; comparison against Historic Environment Record and geophysical survey polygons to derive false-positive and false-negative rates | PDF calibration note with confusion matrix; corrected probability surface with documented accuracy statistics |
| Annual archive-update and attrition monitoring | Incremental addition of each new growing season's acquisitions to the cumulative stack; change detection between successive annual surfaces to identify zones of declining spectral expression | Annual updated GeoTIFF and change-detection layer; alert flag for zones crossing significance threshold |
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.