S.A.B. Kim
INTELLIGENT INTERPOLATION METHODS FOR A FULL-SCALE SPOT-DEM IN COASTAL
REGIONS
S.A.B. KIM', Wonkyu PARK, Tag-Gon KIM"
"Satellite Technology Research Center, " Dept. of Electrical Engineering,
Korea Advanced Institute of Science and Technology, 373-1 Kusung, Yusung, Taejeon, S. Korea 305-701
sbkim@krsc.kaist.ac.kr
KEY WORDS: DEM interpolation, coast, optimal interpolation, quality control, segmentation
ABSTRACT
Intelligent schemes for interpolating stereo-match results to generate a DEM (digital elevation model) are implemented,
as a part of an automatic DEM production from a stereo-pair of satellite images. The need for these schemes is that
interpolation itself is bound to introduce severe blunders, however sophisticated it may be. Due to these blunders,
elevation becomes non-zero over the sea and rivers. Also topography is wrongly created outside stereo-scenes. Even
well-known commercial softwares produce such blunders. To resolve these problems, intelligent schemes are
implemented: firstly, center-of-gravity (COG) and empty-center-index (ECI) which quantify how evenly distributed
interpolants are within an interpolation radius; secondly, hole-fill segmentation scheme to discern whether or not
interpolation should take place in an empty segment of stereo-match results; and last, noise-remove segmentation
scheme to remove noise-like features. The efficacy of these schemes is demonstrated on SPOT DEMs over two 60 km
X 60 km areas in S. Korea. These methods remove all the blunders which amount to 1596 of the stereo-matched points.
1 INTRODUCTION
A digital elevation model (DEM) is one of the most widely used data for terrain analysis and geographical information
systems. DEMs from satellite images have gained increasing advantages over aerial DEMs thanks to regular scanning,
unrestricted access to the global terrain, and high-resolution (up to 1-m) missions. In DEM generation, elevation values
after sensor modeling do not provide complete spatial coverage. Complete coverage may be obtained by interpolating
the scattered elevation values. In this way interpolation becomes a crucial element determining coverage and accuracy
of a DEM.
One of essential issues in interpolation is to find an optimal interpolation method. Renka (1988), Carlson and Foley
(1991), Desmet (1997) and, as a survey paper, Franke (1982) experiments using a test set with less than 100 scattered
input. It was Kim et al. (1999) who study interpolation with a realistic size of input. Firstly, they use an area of 30 km
by 40 km by creating scattered data through random sampling of a 1:25,000 ground survey map. 8 interpolation
methods are applied to the scattered data, and Kriging method shows the best performance. Secondly, they use a SPOT
DEM over a 10 km by 10 km region (1000 by 1000 pixels). Gaussian method shows superior performance to Kriging
thanks to its capability to smooth out stereo-matching errors.
Kim et al. note further that simple application of Gaussian does not produce a satisfactory DEM. In detail, along
the boundaries of a scene, artifacts are generated to a significant degree; such artifacts inevitably occur along the coast;
the elevation over rivers and lakes is interpolated to the elevation of surrounding land. These blunders are exemplified
in Fig. 1. In Fig. 1a one-side of an image is the coast, and the number of such blunders reaches 23 % of the total
matches. Kim er al. introduce preliminary solutions: COG (center-of-gravity) and ECI (empty-center-index)
thresholding.
In this paper, the preliminary solutions are refined and additional schemes for enhancing interpolation performance
are implemented. The series of the enhanced interpolation schemes are named ‘intelligent interpolation’. The need of
intelligent interpolation also exists in state-of-the-art commercial softwares. In this study OrthoEngine v6.2 of PCI and
ImageStation of Intergraph Co. are used. The PCI DEM is displayed in Fig. 1b, where again the elevation along the
coast incorrectly reaches several hundred meters and unrealistic islands are present. To our conjecture interpolation,
together with a ‘pyramid’ approach for stereo-matching, would be the cause of the problems in the PCI DEM. This
paper presents development of intelligent interpolation methods with a view to resolve the limitations in the currently
available DEMs. Section 2 describes data used. In sections 3 and 4, the intelligent interpolation methods and their
performances are presented.
496 International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B3. Amsterdam 2000.
