A study of reducing abnormal elevations in automatic computation of elevations from satellite data

Elevations of the earth surface can be estimated using a pair of stereo images obtained from satellites. The elevations can be automatically computed by an automatic search of corresponding points between the stereo images. Mismatches in searches in search of the corresponding points cause abnormal elevations. In this study, causes of mismatches were investigated in computing elevations using Landsat TM data. The variance of source window and the correlation coefficient at matched point were investigated by a discriminant analysis between the two groups of abnormal and normal elevation points. The discriminant analysis showed that the abnormal elevation points could be discriminated from normal elevation points by examining the variance of the source window and the correlation coefficient at matched points with a pretty high probability of about 80%. A method of computing elevations for reducing abnormal elevations points using the following two ideas was proposed: - Enlarge the size of source window if the variance of the source window is smaller than a threshold. - Assign a specified elevation if the correlation coefficient at matched points is smaller than a threshold and compute a proper elevation by an interpolation in post-processing. An elevation image generated using the above method from Landsat TM data shows the method is effective for reducing abnormal elevations.