Abstract
Information on terrain features like slope, orientation and convexity may be very useful for thematic interpretation of single band satellite radar data. Accuracy of the co-registration is the key issue. The map-projected GBFM radar mosaic of Siberia has been co-registered with a digital elevation model, using a cross-correlation technique in the Fourier domain. The mosaic was produced in the framework of the Global Boreal Forest Mapping Project, an initiative of the Japan Agency for Space Exploration (JAXA), and is based on JERS-1 synthetic aperture radar (SAR) data acquired in years 1997-8. The SRTM digital elevation data (90 m horizontal resolution) have been used for areas up to 60 degrees of latitude and the USGS GTOPO30 elevation data (500 m horizontal resolution) for the rest of the area. Since SAR and DEM data-sets capture completely different features of the landscape and SAR imagery is affected by geometric and radiometric (shadow and layover) distortions due to elevation and local terrain slope, automatic matching by homologous features of the radar image to the DEM image is not possible. Due to the unavailability of SRTM data at the time of the mosaic processing and, in any event, due to computational constraints (the mosaic is composed of some 400 SAR strip-images covering 135 000 km2 each) the classical geo-coding procedure using slant range data had to be ruled out. The a-posteriori solution entailed the simulation of the radar reflectivity dependency on the local incidence angle based on available DEM and radar viewing geometry. The radar mosaic was then matched with the simulated image. The cross-correlation moving window was composed of mutually overlapping squares (60 by 60 pixels) in a regular grid with 20 pixel spacing between the centers. The co-registration gives good level of correlation not only for mountainous areas but also for hilly ones. High correlation occurs also in flat areas with pronounced hydrological features like river courses and lake shores that are reflected in SRTM as fine-detail features. The density of control points was on average 1400 points per 100 square kilometers. The geometric effect of topography (like shortening of the slopes oriented towards the radar) has been then corrected by inversion of the same model which had been used for generating the simulated image. Radar backscattering coefficient dependency on local incidence angle was modeled and corrected by a simple inverse sine function model. The corrected radar image was then fed to a classification algorithm together with layers extracted from the DEM, such as slope and convexity. Preliminary thematic classification results confirm that geometric and radiometric corrections afforded by this technique greatly improve the classification accuracy.
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have