Introduction

Loosely defined, the process of stratigraphic interpretation is the analysis of the dip distribution (or dip spectrum) of a seismic image in small neighborhoods and the corresponding association of local geology with a given stratigraphic sequence. The interpreter's job, illustrated in Figure 1, is both tedious and time-consuming if performed manually, considering the large size of modern 3-D surveys. For instance, if a given sedimentary unit is best defined by its relative distance from a pervasive geologic unconformity, the interpreter must first identify the location of the unconformity over the entire seismic volume.

 

algorithm
Figure 1 Stratigraphic interpretation: Given a facies template (left), search a seismic image locally (center) for likely matches to the template, then output (right) an attribute which illustrates some measure of local similarity between the template and the seismic image.

To make the stratigraphic interpretation of large 3-D image volumes feasible, an automatic approach is required to search an image locally for the likely presence of a predefined ordered pattern, or facies template. Randen et al. (1998) presented an automated scheme which analyzes local dip spectra to detect reflector terminations in seismic images, and hint that such an approach could be used to detect unconformities and recognize facies patterns. Neural networks have been applied to this end, but the results are often non-intuitive.

We present a scheme which automatically searches a seismic image for an arbitrary facies template, and then outputs a similarity attribute which expresses the data's relative local resemblance to the template. To compute the similarity attribute, we recast this problem of pattern recognition to one of signal/noise separation, i.e., treating the facies template as the ``noise model'', we seek to remove an optimal amount of it from small data windows, then define the attribute as the local noise-to-signal ratio. It follows that the similarity attribute is both physically meaningful and optimal in one (least squares) sense.

We first test the scheme on a 2-D synthetic seismic image with two unconformities, and find that both are detected reliably. We then perform the same test on a 2-D real seismic image, and successfully detect an unconformity. The performance of the scheme is encouraging, and there is considerable room for optimization.

 

algorithm2
Figure 2 For each output point, extract a neighboring window of data the same size as the facies template. Capture the local dip spectra of the facies template and data window with simple nonstationary ``steering filters''. Treating the facies template as the ``noise'' model, apply a predictive signal/noise separation technique to extract energy from the data window where the local dip coincides with that of the template. The output, simply the noise-to-signal ratio, is then a valid measure of local similarity.