Automated Geometric Correction of High-Resolution Pushbroom Satellite Data

In this article, we present the use of the Automatic Ground control points Extraction technique (AGE) for increasing the automation in the geometric correction of high-resolution satellite imagery. The method is based on an image-to-image matching between the satellite data and an already geocoded image (i.e., a digital orthophoto). By using an adaptive least squares matching algorithm which implements a very robust outlier rejection technique, AGE can automatically measure many hundreds of topographic features (TFs) on the images, whose cartographic coordinates are derived from the geocoded image and elevations are extracted from an associated digital elevation model (DEM). The AGE technique has been tested for different high-resolution data: (a) 0.62-meter QuickBird panchromatic data (basic imagery processing level), (b) 2.5-meter SPOT-5/HRG panchromatic supermode data (standard 1B processing level), and (c) 1-meter Ikonos panchromatic data (standard Geo product processing level) collected in the Northern of Italy, both in flat (Torino Caselle test site) and mountain areas (Lecco test site). Regardless the relative image resolution between the satellite and the aerial data (1-meter) and regardless the processing level of the original satellite data, a similar TFs density has been obtained for both the QuickBird and the SPOT-5/HRG data (4.4 GCPs/km2 and 4.1 GCPs/km2) respectively, with a geometric accuracy for the GCPs extracted of 0.90 m for QuickBird and 3.90 m for SPOT-5/HRG. For the Ikonos imagery, AGE extracted a more dense set of GCPs (8.7 GCPs/km2) but with a lower accuracy (3.19 m). The TFs identified with AGE can be used as GCPs for the rational polynomial coefficients (RPCs) computation and, therefore, for implementing a full automatic orthoimage generation procedure. By using the commercial off-the-shelf software PCI Geomatica® v.9.1, orthoimages have been generated for all datasets. The geometric accuracy was verified on a set of 30 manually measured independent check points (CPs) and assessed a precision of 4.99 m RMSE for QuickBird, 5.99 m RMSE for SPOT-5/HRG, and 8.65 m RMSE for Ikonos. The use of a non-conventional image orthorectification technique implementing a neural network GCPs regularization, tested for the SPOT-5/HRG data, showed the full potential of the AGE method, allowing to obtain a 3.83 m RMSE orthoprojection in a fully automated way.