Geometric Correction, Registration, and Resampling of Landsat Imagery

Abstract

The geometric distortions introduced by the LANDSAT imaging system are well known. However, the satellite does not provide attitude information with sufficient accuracy to determine the geographical position of each picture element or pixel. Thus, to map the data onto the desired space it is necessary to model a geometric transformation from a set of ground control points. Transformation of a LANDSAT image to a UTM co-ordinate grid to 50 metres accuracy (rms) under optimum conditions has been demonstrated. The main factor limiting the ultimate accuracy of the corrected image is the precision to which ground control points on the UTM maps can be mapped to pixels in the LANDSAT image. Digital correlation methods have been used successfully in image-to-image registration to improve the accuracy by 25%, but for map registration, this requires accurately located ground control photographic chips. Various resampling techniques including sin (Πx)/Πx expansions, cubic splines, and Lagrange polynomials are described. The errors introduced by these resampling methods are investigated both theoretically and experimentally.