IMAGE FILTERING BY FACTORIAL KRIGING - SENSITIVITY ANALYSIS AND APPLICATION TO GLORIA SIDE-SCAN SONAR IMAGES

Factorial Kriging (FK) is a data-dependent spatial filtering method th at can be used to remove both independent and correlated noise on geol ogical images as well as to enhance lineaments for subsequent geologic al interpretation. The spatial variability of signal, noise, and linea ments, characterized by a variogram model, have been used explicitly i n calculating FK filter coefficients that are equivalent to the krigin g weighting coefficients. This is in contrast to the conventional spat ial filtering method by predefined, data-independent filters, such as Gaussian and Sobel filters. The geostatistically optimal FK filter coe fficients, however, do not guarantee an optimal filtering effect, if f ilter geometry (size and shape) are not properly selected. The selecti on of filter geometry has been investigated by examining the sensitivi ty of the FK filter coefficients to changes in filter size as well as variogram characteristics, such as nugget effect, type, range of influ ence, and anisotropy. The efficiency of data-dependent FK filtering re lative to data-independent spatial filters has been evaluated through simulated stochastic images by two examples. In the first example, bot h FK and data-independent filters are used to remove white noise in si mulated images. FK filtering results in a less blurring effect than th e data-independent filters, even for a filter size as large as 9 x 9. In the second example, FK and data-independent filters are compared re lative to the extraction of lineaments and components showing anisotro pic variability. Ir was determined that square windows of the filter m ask are effective only for removing isotropic components or white nois e. A nonsquare windows must be used if anisotropic components are to b e filtered out. FK filtering for lineament enhancement is shown to be resistant to image noise, whereas data-independent filters are sensiti ve to the presence of noise. We also have applied the FK filtering to the GLORIA side-scan sonar image from the Gulf of Mexico, illustrating that FK is superior to the data-independent filters in removing noise and enhancing lineaments. The case study also demonstrate that variog ram analysis and FK filtering can be used for large images if a spectr al analysis and optimal filter design in the frequency domain is prohi bitive because of a large memory requirement.