Digital image interpretation by analysis of image-image correspondence

Abstract

The correspondence of uninterpreted digital images derived from a space-borne multi-spectral scanner (MSS) with high resolution images was studied in developing a quantitative approach to digital image interpretation. In general terms, the method involved enumeration of the co-occurrence of interpreted classes from a familiar image with the unknown classes of a digital image and analysis of the observations organized as a matrix. Specifically, the technique was applied to projects in which the digital images derived from LANDSAT MSS data were interpreted. Target information classes were delineated on large-scale photography, and registration of the digital and interpreted photographic images permitted the assignment of two characterizations for any ground location, one being the digital class derived from MSS data and the other being the target class. For each subsequent analysis, counts of the frequency of co-occurrence of all observed digital and target class combinations were organized as a matrix, X. The number of times a given target class, i, occurred with a given digital class, j, represented an observation, x j , in X. Prior to analysis, matrices were condensed to reduce redundancy and cleaned to reduce noise. Analyses applied included calculations for Yule's coefficient for target-digital class association, chi-square tests of association, principal components analysis, correspondence analysis, and nonmetric multi-dimensional scaling. Chi-square test results were consistently useful and were used to transform matrices of observational data to a binary form which summarized target-digital class associations. Yule's coefficient was useful only in comparing classes of equivalent frequency. Inconsistent results were obtained when the other multivariate techniques were used with observational data, but with data transformed by the chi-square test results were stable and interpretable. In two of three applications of correspondence methods, about two percent of the total digital image area was sampled with large-scale photography, and in the third application enumeration of class-class co-occurrence was complete. A comparison of results suggests that sampling intensities of one to two percent would be adequate for many applications and that higher sampling intensities might not be economic. Two of the three applications were in areas of high relief, and registration of digital and photographic images was by bilinear interpolation of control points. This induced considerable noise in observational data due to effective misregistration and suggests higher order methodologies should be considered in image registration in such areas. The third application also involved bilinear interpolation, but was in an area of minimal relief, and there was little apparent noise. This approach to interpretation of digital images was developed empirically and has been employed in interpretation of imagery only in conjunction with other traditional interpretive approaches. Results to date do, however, invite consideration of employing this approach, after research and refinement, singly in such tasks as automated target detection and photointerpretation and empirical evaluation of sensor discrimination capability.