Merging of airborne elevation data and Radarsat data to develop a Digital Elevation Model

Abstract

A Digital Elevation Model (DEM) has been developed for a site in French Guiana by merging airborne elevation data with a stereo‐radargrammetric DEM. Two Synthetic Aperture Radar (SAR) images were available from the Radarsat sensor acquired in S2 standard and F5 fine modes, as well as airborne laser and radar data distributed along flight lines with a 500‐m spacing giving a data interval of every 7 m for the laser data and every 70 m for the radar data. The radargrammetric DEM was extracted using a regular 50 m×50 m grid pattern from the Radarsat images and its accuracy estimated. Pre‐processing of the airborne data enabled the anomalous points to be discarded and the top‐of‐canopy points to be selected. Kriging was then applied to merge elevations from the DEM, not very accurate but dense, with those from the airborne altimeters, accurate but irregular and sparse. Two types of kriging were performed: (1) kriging of the difference between DEM elevations and airborne data, which was then subtracted from the radargrammetric DEM, and (2) direct kriging of airborne data when radargrammetric DEM elevations were unsatisfactory. Merging significantly improved the accuracy of the radargrammetric DEM. The standard deviation of elevation errors fell from 21.2 m to 13.8 m or from 25.3 m to 11.0 m, depending on the check source adopted.

Acknowledgments

The present study was jointly funded by BRGM (Bureau de Recherche Géologiques et Minières) and CNES (Centre National des Etudes Spatiales). The Radarsat data were provided free of charge as part of the ADRO2 programme of the Canadian Space Agency. The heliborne DEM used was built by the company ALTOA on behalf of the mining company ASARCO. The authors would like to thank Rowena Stead for editing the English, and also the two anonymous reviewers for their time spent reviewing and improving this paper.

Notes

Exponential model:

where a=scale factor, 3a=practical range, C = sill, and h=interdistance.

Linear model:

where p=model's slope, α=1 (in our case).