Automatic identification of ground control points in synthetic aperture radar images

Ground control points (GCPs) are points on the Earth's surface that can be used to geo-reference remote sensing satellite data. The identification of GCPs can be achieved either by on-ground inspections or by processing of aerial/satellite data. The first option ensures the highest geolocation accuracy, but it can require long turnaround times and high logistical costs. Moreover, in situ inspections can be impossible in certain areas of the world (e.g., because of difficult access or geopolitical issues). Hence, there is a strategic interest in the development of services that can be used to deliver accurate remote sensing of GCPs. Ongoing satellite missions—such as the Constellation of Small Satellites for Mediterranean Basin Observation (COSMO-SkyMed, or CSK) and TerraSAR-X (TSX)—are providing high-resolution synthetic aperture radar (SAR) data in the X-band (i.e., a segment of the microwave radio spectrum used for communications). The availability of this data is fostering the development of methods and services for extracting GCPs with sub-metric accuracy. Such endeavors are also appealing for both civilian and military purposes.

Conceptually, SAR-based GCP-extraction consists of two steps. First, the same local feature is identified in a number of SAR images and its range/azimuth coordinates are determined. The second step involves the use of spatial triangulation (stereo analysis) and inversion methods for spatial 3D position retrieval from the 2D radar coordinates (range/azimuth). To boost the geolocation accuracy, SAR images must be acquired from different lines of sight, with intersection angles typically much wider than 10°. The lines of sight can even be in completely opposite directions,¹ but this usually hinders the ability to perform the GCP identification in a fully automatic manner. In addition, many robust algorithms—such as the Scale-Invariant Feature Transform (SIFT), Speeded Up Robust Features, and Binary Robust Independent Elementary Features methods—have been proposed for the automatic identification and matching of local invariant features.² These algorithms, however, were originally developed for optical photographs and images, and do not perform well on images affected by multiplicative noise sources (e.g., speckle in SAR data). To mitigate the effect of speckle noise, many adaptations of local-features-matching algorithms have thus been proposed in the recent literature, including SAR-SIFT³ and SIFT-Octave.⁴ These algorithms, however, only have satisfactory performances with narrow intersection angles.

In this study,⁵ we propose a further adaptation to the existing algorithms. Our approach is specifically designed to ensure robustness and accuracy in the fully automatic detection of bright isolated targets (e.g., steel light poles or towers), even when dealing with opposite-side-looking data. In particular, we have selected the popular Harris algorithm⁶ as the detector. We chose this algorithm because recent studies^{4, 7} provide a consensus that it is the most stable and robust-to-noise algorithm for corner detections from SAR images. Furthermore, we opted for a solution that combines simplicity and robustness. This is different from previously proposed descriptors, which typically consist of high-dimension feature vectors (e.g., 128 for SIFT).

We derive the properties of our descriptor from radar fragments centered on the candidate GCP corner. We also perform mild geocoding, through the use of low-resolution digital terrain models (DTMs) that are available over the area of interest (e.g., the Shuttle Radar Topography Mission or radargrammetric DTMs). In addition, the performance of our matching algorithm can be enhanced by limiting the search space (as demonstrated by previously proposed solutions, in which DTM and orbit information are used⁷ or that involve a manual pre-registration⁸). Once we have identified the GCP candidates and have estimated their fractional line/pixel coordinates from each SAR image, it is necessary to convert those coordinates into sensor–target distances and azimuth coordinates.

In our approach, we must also account for the atmospheric propagation delay (APD)—i.e., delays in satellite-to-Earth communications caused by the atmosphere—for computation of the slanted range distance. We thus use numerical weather modeling⁹ to precisely estimate the APD. In the final part of our technique, we invert range/Doppler equations to infer the 3D geographic coordinates of the candidate GCPs. To improve the accuracy and confidence level, we use multiple scenes (with different look-angles) and we use the least-squares method to solve the over-determined system of equations.

Although automatic GCP-extraction tools have previously been successfully verified on TSX data,¹⁰ such methods have not yet been tested on CSK images. We have therefore focused on assessing the performance of our algorithm for the processing of CSK data. According to the results from our earlier study,¹¹ a preliminary timing recalibration is required for CSK images. This step is necessary for removing systematic geolocation biases in both range and azimuth directions. In this work, we conducted our experimental analysis for the Pisa area (Italy), for which consistent data takes of radargrammetric enhanced spotlight images are available in the CSK archive. An example of a GCP, in the Pisa urban area, which we automatically detected with our algorithm is shown in Figure 1. The outcomes of our analyses prove that our designed tool is capable of automatically identifying GCPs from CSK data and estimating their geographic position with sub-metric accuracy. In addition, as part of a differential global positioning system campaign, we have computed geolocation errors for a number of vertical isolated targets.

Figure 1. Example of an automatic detection of a ground control point (GCP) in the urban area of Pisa (Italy). Arrows indicate the position of the detected light tower (i.e., the GCP), which is installed within the Don Mario Azzola traffic circle, in a (top) synthetic aperture radar image from the Constellation of Small Satellites for Mediterranean Basin Observation, (inset) a photograph, and (bottom) an image from GoogleEarth.

In summary, we have developed a new algorithm for the automatic detection of ground control points in remotely sensed SAR datasets. We have based our tool on the popular Harris algorithm, and have thus designed a simple, robust, and accurate algorithm. We have also successfully tested our approach on data from the COSMO-SkyMed satellite for an urban area and have verified that it can be used to estimate the geolocation of GCPs, with sub-metric accuracy. In the next stage of our work we will assess the performance of the tool over mountainous areas, i.e., where radar image distortions are expected to hinder the visibility of GCPs (depending on the orientation of the slopes and the viewing geometry of the satellite sensors).

This study was carried out within the framework of the 3D IMINT project (PRNM Contract 10444, 30-12-2013). CSK products, under a license from the Italian Space Agency, were used for this project.