SAR and optical fusion for systematic tells and burial mound mapping
Archaeological tells and burial mounds produce distinctive SAR backscatter asymmetries and DEM slope signatures that separate them from natural hillocks. Fusing Sentinel-1, ALOS-2 and high-resolution optical data in machine-learning classifiers makes systematic survey across thousands of square kilometres practical, though field validation remains non-negotiable.
Sensors
- Sentinel-1 C-band SAR (ESA): 5.6 cm wavelength; IW mode ground range resolution approximately 5 x 20 m, resampled to 10 m; 6-day repeat at mid-latitudes. Provides amplitude backscatter and shadow-layover geometry sensitive to mound flanks. Free and openly archived from 2014.
- ALOS-2 PALSAR-2 L-band SAR (JAXA): 23.6 cm wavelength; Fine Beam Single mode at 3 m range resolution. L-band penetrates sparse vegetation better than C-band, improving detection of low mounds obscured by scrub. Revisit approximately 14 days; commercial tasking required for most areas.
- Copernicus DEM GLO-30: Global 1 arc-second (approximately 30 m posting) DEM derived from TanDEM-X radar interferometry. Absolute vertical accuracy better than 4 m LE90 over most terrain. Sufficient to detect tells above roughly 3-5 m height; misses low kurgans in flat steppe without supplementary data.
- WorldView-2 / WorldView-3 (Maxar): Panchromatic resolution 0.31-0.46 m; 8-band multispectral at 1.24-1.85 m. Used for texture classification and shadow-length photogrammetry of individual mounds. Tasking costs apply; not freely archived.
What a mound looks like to a radar
A tell or burial mound is, from a SAR perspective, a small symmetric hill with a characteristic foreshortening-shadow pair. The illuminated flank facing the sensor returns high backscatter; the opposite flank falls into radar shadow. The azimuthal orientation of that shadow, combined with the local incidence angle, encodes the mound's height-to-diameter ratio. Natural erosional hillocks produce similar signatures, but their planimetric shape is rarely as circular or elliptical as a constructed mound, and their slope profiles differ in measurable ways.
C-band Sentinel-1 at 10 m resolution can resolve tells above roughly 20-30 m basal diameter, which covers most major Bronze Age and earlier mounds in the Near East and Eurasian steppe. Smaller kurgans, some only 10-15 m across, require the 3 m resolution of PALSAR-2 Fine Beam or commercial SAR such as ICEYE or Capella. Published work on the Syrian Jazira, including systematic surveys by Menze and Ur (2012) using CORONA and SRTM, established that several thousand tells exist in that region alone, the majority unexcavated and many unregistered.
Fusion architecture: why one sensor is not enough
SAR amplitude alone produces an unacceptably high false-positive rate when applied across mixed terrain. Rock outcrops, irrigation berms, road embankments and even dense reed patches all mimic mound backscatter. The standard mitigation is to fuse SAR-derived slope and shadow features with optical texture layers and a high-resolution DEM in a supervised classifier.
The typical pipeline extracts from SAR: local incidence-angle-corrected backscatter, the ratio between ascending and descending passes (which amplifies symmetric features), and the shadow-layover mask. From the optical layer it extracts grey-level co-occurrence matrix (GLCM) texture features, which distinguish the compacted, often vegetation-sparse surface of an earthen mound from surrounding agricultural or natural ground. The Copernicus GLO-30 DEM contributes slope, curvature and topographic position index. A random-forest or gradient-boosted classifier trained on a labelled set of confirmed mounds and confirmed non-mounds then assigns probability scores across the survey area. Precision-recall trade-offs are explicit: a low detection threshold recovers more true mounds but multiplies the field-validation burden.
What the published record shows, and where it falls short
Systematic remote-sensing surveys of the Syrian Jazira documented over 14,000 potential tell sites using CORONA declassified imagery and SRTM topography. More recent work applying machine-learning classifiers to Sentinel-1 and optical data in the Eurasian steppe has reported precision figures in the range of 60-80 per cent and recall figures of 55-75 per cent, depending heavily on training-set quality and terrain homogeneity. Those are honest numbers: one in four to one in five detections is a false positive, and a meaningful fraction of real mounds are missed.
Cloud cover is a practical constraint only for the optical component. SAR acquisition is weather-independent, which matters in the Caucasus and parts of the steppe where optical windows are limited seasonally. The GLO-30 DEM is static; it cannot detect mounds that have been partially levelled by agriculture since the TanDEM-X acquisition period (2010-2015), nor can it resolve mounds below its vertical noise floor of roughly 2-4 m. For low-relief kurgans on flat steppe, lidar or stereo-photogrammetric DSMs at sub-metre posting are the only reliable topographic input, and no open global product at that resolution currently exists.
The validation burden that remote sensing cannot remove
Every candidate detection list produced by a classifier is a hypothesis, not a site register. Ground-truthing or, where access is denied, very high resolution optical confirmation at sub-0.5 m is required before any candidate enters a heritage inventory. In conflict-affected areas such as northern Syria or parts of Ukraine, ground access may be impossible for years. In those cases, WorldView-3 panchromatic imagery at 0.31 m can resolve surface morphology well enough to confirm or reject most candidates, but it cannot substitute for the stratigraphic information that defines a site's cultural period and significance.
The validation burden scales with survey area. A 50,000 km² survey producing 2,000 candidates at 70 per cent precision leaves roughly 600 false positives to be resolved. Prioritisation algorithms, ranking candidates by morphometric confidence score and proximity to known sites, can concentrate fieldwork efficiently, but the resource implication for heritage agencies must be stated clearly at project inception.
Practical scope and what this workflow is actually good for
The strongest application is rapid triage across large areas where no systematic field survey has ever been conducted, or where pre-existing registers are decades old. A government heritage agency or international programme can use a classifier-derived candidate layer to plan field seasons, prioritise conservation investment and identify which mounds face imminent agricultural or development threat. That is a genuine capability advance over manual interpretation of aerial photographs.
A secondary application is change detection. By running the same classifier over multitemporal SAR stacks, it is possible to flag mounds whose backscatter signature has changed, suggesting levelling, excavation or erosion. This is distinct from looting-pit detection (covered in a sibling page) and focuses on gross morphological change at the mound scale. Satellize runs this kind of multitemporal fusion analysis on open Sentinel-1 archives, and the same analytical framework that underpins the Tonga crop-estimation programme, systematic classification of surface features across large areas using open satellite data, transfers directly to archaeological candidate mapping.
The method works best in semi-arid and arid environments with sparse vegetation, moderate topographic relief and limited urban sprawl. Dense forest, high-relief mountain terrain and heavily irrigated lowlands all degrade classifier performance significantly.
Specifying a survey: questions to answer before commissioning data
Before any data order is placed, four parameters need to be fixed. First, the minimum target size: if the priority is kurgans under 15 m diameter, Sentinel-1 is insufficient and PALSAR-2 or commercial SAR must be budgeted. Second, the acceptable false-positive rate and the field-validation capacity available to resolve it. Third, whether the survey area has existing labelled mound inventories large enough to train a supervised classifier, or whether transfer learning from a geographically similar region is necessary. Fourth, the archive depth required: Sentinel-1 data from 2014 onwards is free; PALSAR-2 archive access is negotiated separately with JAXA.
These are not bureaucratic questions. Each one changes the cost, timeline and reliability of the output. A classifier trained on 50 labelled mounds in a different country will underperform one trained on 500 local examples. That gap is recoverable with additional labelling effort, but not with additional satellite data alone.
Typical figures
| SAR spatial resolution (Sentinel-1 IW) | ~10 m (resampled); range resolution ~5 m, azimuth ~20 m native |
| SAR spatial resolution (ALOS-2 PALSAR-2 Fine Beam) | ~3 m range; suitable for mounds ≥10 m basal diameter |
| DEM vertical accuracy (Copernicus GLO-30) | <4 m LE90 over most terrain; mounds below ~3-5 m height unreliable |
| Optical resolution (WorldView-2/3 panchromatic) | 0.31-0.46 m; enables shadow-length height estimation and surface confirmation |
| Sentinel-1 revisit (mid-latitudes) | 6 days (combined ascending + descending); 12 days single pass |
| Sentinel-1 archive depth | 2014 to present; free via Copernicus Data Space |
| Minimum detectable mound (Sentinel-1 fusion) | ~20-30 m basal diameter, >2 m height; smaller targets require commercial SAR |
| Classifier precision/recall (published steppe surveys) | 60-80% precision, 55-75% recall; highly training-set dependent |
| Coverage per processing run | Scalable to 100,000+ km² per Sentinel-1 scene mosaic; no practical upper limit on open data |
| Delivery formats | GeoPackage or shapefile of candidate polygons with probability scores; GeoTIFF probability raster; PDF morphometric report per candidate class |
Analytics Satellize can run
| Mound candidate layer with probability scores | Random-forest or gradient-boosted classifier trained on SAR amplitude, ascending/descending backscatter ratio, GLO-30 slope and curvature, and GLCM optical texture features | GeoPackage of candidate polygons with per-feature confidence score, morphometric attributes and false-colour composite thumbnails |
| Morphometric attribute table | Automated extraction of basal diameter, height estimate (from DEM or shadow length), circularity index and slope symmetry for each candidate | CSV or geodatabase table joinable to candidate layer; sortable by confidence rank for field-season planning |
| Change-detection alert layer | Multitemporal SAR backscatter differencing on Sentinel-1 annual composites to flag gross morphological change at mound scale | GIS layer of mounds showing statistically significant backscatter change, with date range and magnitude; suitable for conservation priority triage |
| False-positive triage report | Secondary classifier stage using WorldView panchromatic texture and planimetric shape metrics to separate anthropogenic mounds from natural hillocks and infrastructure | Ranked shortlist of highest-confidence candidates with sub-metre optical chip per site; formatted for heritage agency review |
| Survey-area coverage statistics | Systematic grid-based completeness assessment: proportion of survey area with valid SAR acquisition, cloud-free optical coverage and DEM data within specification | Coverage gap map in GeoTIFF; identifies zones requiring additional tasking before classifier results are reliable |
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.