Geometric accuracy testing of satellite remote sensing imagery verifies the positional fidelity of spaceborne image products through rigorous comparison against high-precision ground reference data. This article explains the core accuracy indicators—such as CE90, RMSE, internal geometric distortion, and band-to-band registration—along with applicable standards, sampling strategies, and evaluation procedures used by accredited third-party laboratories for optical and SAR satellite imagery products.
Technical Overview
Geometric accuracy testing (geometric precision verification) of satellite remote sensing imagery is a quantitative evaluation process that measures how faithfully the positions of features in an image product correspond to their true locations on the Earth's surface. It is a mandatory quality control step in the lifecycle of satellite remote sensing products, covering in-orbit commissioning of new sensors, routine production of standard image products, and acceptance inspection before imagery is delivered for surveying and mapping applications.
The evaluation addresses several distinct aspects of geometric quality: absolute geolocation accuracy, which describes the positional offset of the image in a defined geodetic reference frame without the use of ground control points; relative or internal geometric accuracy, which describes distortions within a single scene after removing the best-fitting systematic transformation; band-to-band registration accuracy for multispectral and fusion products; and height accuracy for stereo products derived from along-track or cross-track imaging. Results are expressed through statistical indicators such as root mean square error (RMSE), circular error at 90% or 95% probability (CE90/CE95), linear error (LE90/LE95), and normalized median absolute deviation (NMAD).
The work is typically performed by accredited third-party laboratories in accordance with international and national standards, including the ISO 19159 series on calibration and validation of remote sensing imagery sensors, ISO 19157 on geographic information data quality, and Chinese national standards such as GB/T 18316 and GB/T 24356 for quality inspection and acceptance of digital surveying and mapping products. The evaluation framework also aligns with the quality assurance principles promoted by the Committee on Earth Observation Satellites (CEOS) under its QA4EO guidelines, which require that all delivered data products carry documented and traceable quality indicators.
Testing Samples
Level-0 to Level-1 Raw and Systematic Corrected Products
These include radiometrically corrected imagery with systematic geometric correction applied using onboard orbit and attitude data, commonly designated as Level-1B or Level-1C products. Testing at this level characterizes the direct geolocation performance of the satellite platform and sensor geometry without ground control.
Orthorectified Level-2 Products
Geo-corrected orthophoto products generated using rational polynomial coefficients (RPC), ground control points, and digital elevation models. Testing verifies whether the orthorectification process achieves the specified positional accuracy for the target mapping scale.
Multispectral and Panchromatic Band Sets
Multiband image sets, including panchromatic and multispectral bundles intended for pan-sharpening, are tested for band-to-band and pan-to-multispectral registration accuracy, which directly affects the quality of fused products.
Stereo Image Pairs and SAR Products
Along-track or cross-track stereo pairs are tested for planimetric and elevation accuracy through space intersection. Spaceborne SAR imagery, including geocoded products, is evaluated for geolocation accuracy in slant or ground range geometry.
Sample Submission Requirements
- Complete image files with associated metadata, RPC files, and projection definitions
- Product level description, processing date, and sensor mode information
- Imagery preferably covering terrain with accessible, well-defined ground features
- Cloud coverage generally below 20% over the verification area
Testing Items
- Absolute geolocation accuracy: expressed as RMSE in easting and northing, radial RMSE, or CE90/CE95 circular error, evaluated without ground control points
- Relative (internal) geometric accuracy: residual distortion within the scene after fitting an affine or polynomial transformation, reflecting detector alignment, push-broom stability, and attitude jitter effects
- Band-to-band registration accuracy: sub-pixel offset between spectral bands measured by cross-correlation of corresponding image chips
- Pan-sharpened fusion product geometric consistency between panchromatic and multispectral layers
- Stereo positioning accuracy: planimetric accuracy and height accuracy (LE90 or RMSE of elevation residuals) from space intersection of stereo pairs
- Orthorectified product positional accuracy against reference orthophoto basemaps or field-surveyed checkpoints
- Geometric consistency across detector stitching lines, across-track gaps, and temporal composites
- Pixel resolution and grid conformity: verification of ground sample distance and geodetic grid definition against product specifications
Testing Methods
Field Checkpoint Method
High-accuracy ground checkpoints are established in the image coverage area using static GNSS surveying or network RTK techniques. Checkpoints are selected on stable, unambiguously identifiable features such as road intersections, field corners, and marking points. The reference point accuracy should be at least three times better than the expected image accuracy, in accordance with the tolerance requirements of GB/T 18316. Image coordinates of each checkpoint are measured, residuals against the reference coordinates are computed, and RMSE and CE90 values are derived statistically.
Reference Imagery Matching Method
Instead of field surveying, image chips from the product under test are automatically matched to an authoritative reference orthophoto basemap of verified higher accuracy, typically resampled from higher-resolution reference data. Sub-pixel image matching algorithms based on normalized cross-correlation or phase correlation generate dense correspondences, from which systematic offsets and local distortions are quantified. This method enables efficient large-area verification and dense internal accuracy assessment.
Stereo Intersection Method
For stereo products, ground coordinates are computed by space intersection of conjugate points measured on the stereo pair, and the resulting three-dimensional positions are compared with reference checkpoints to evaluate both planimetric and vertical accuracy expressed through CE90 and LE90 indicators.
Band Registration Evaluation
Corresponding image windows from different spectral bands are registered using sub-pixel correlation techniques. The measured offsets in along-track and across-track directions are statistically summarized to verify compliance with the sub-pixel registration specification, commonly required to be better than 0.3 to 0.5 pixel.
Sampling and Statistical Evaluation
- Checkpoints are distributed evenly across the scene using stratified sampling, avoiding water bodies, shadowed areas, and terrain discontinuities
- A minimum of 20 to 25 clearly defined points per scene is generally required, with additional points for large-area products
- Statistics reported include mean error, standard deviation, RMSE, CE90/CE95, LE90, and NMAD for robustness against outliers
- Rejection criteria and pass thresholds follow the product specification and the applicable acceptance standard
Deliverables
The laboratory issues a test report containing the sampling scheme, checkpoint distribution map, residual vector diagram, statistical accuracy indicators, uncertainty statement, and a conformity conclusion against the specified accuracy class. Reports issued under ISO/IEC 17025-accredited quality management systems are widely accepted for procurement acceptance and regulatory submission.
Applications
- In-orbit geometric calibration and commissioning tests of newly launched Earth observation satellites, including validation of sensor alignment and RPC models
- Acceptance inspection of standard image products delivered by satellite ground segments and data suppliers
- Quality verification supporting topographic mapping at scales from 1:2,000 to 1:50,000 and national geographic database updates
- Compliance evaluation for government procurement of high-resolution imagery in natural resources, agriculture, forestry, and ecological monitoring programs
- Baseline accuracy confirmation for change detection, image fusion, and long-term time-series analysis projects
- Support for emergency mapping and disaster response, where documented positional accuracy is critical for reliable damage assessment
- Pre-delivery verification for international data supply, where geometric quality declarations must meet contractual and ISO metadata requirements
FAQ
What is the difference between CE90 and RMSE?
RMSE is the root mean square of coordinate residuals and describes the average magnitude of positional error in one or two dimensions. CE90 is a circular error measure indicating the radius of a circle within which 90% of positional errors are expected to fall, derived from the probability distribution of the radial errors. CE90 is the most widely quoted indicator for satellite imagery geolocation accuracy, while RMSE provides the underlying statistical detail.
What accuracy must the reference data have?
The reference data, whether field-surveyed checkpoints or reference orthophoto basemaps, should be at least three times more accurate than the product under test. For example, verification of imagery specified at 5 m CE90 requires reference points accurate to better than approximately 1.5 m, ensuring that reference uncertainty does not dominate the measured result.
How many check points are required for a reliable evaluation?
Standard practice requires a minimum of 20 to 25 well-defined, evenly distributed points per scene, following the sampling rules of GB/T 18316 and GB/T 24356. Larger mosaics or products with complex terrain require additional stratified samples to characterize accuracy variation across the coverage area.
Which product levels can be tested?
All standard product levels can be evaluated, including systematic corrected Level-1 products for absolute geolocation without control, orthorectified Level-2 products for mapping-grade accuracy, multispectral bundles for band registration, and stereo pairs for elevation accuracy verification.
How long does a geometric accuracy test take?
For standard laboratory evaluation using reference imagery matching, the typical turnaround is 5 to 10 working days per batch of scenes, including matching, statistical analysis, and report preparation. Projects involving field GNSS checkpoint campaigns require additional scheduling for site access and survey execution.
Is the test report accepted for procurement and compliance purposes?
Yes. Reports issued by accredited third-party laboratories operating under ISO/IEC 17025 quality systems, with testing performed in accordance with ISO 19159, ISO 19157, and applicable GB/T standards, are recognized for satellite product acceptance, government procurement, and contractual quality disputes.
