Interpreter input for seismic image segmentation

Interpreter guidance

It is clear from the initial 2D segmentation result in Figure 2(a) that the algorithm did not perform ideally in several locations. Specifically, the salt segment ``leaks'' through the boundary at two locations, one along the top of the salt body and one along the base. In such situations, we want to use interpreter input to rectify the inaccuracies. Figure 2(b) shows manual interpretations of the salt boundary in the region where the leakages occurred. These boundary segments were quickly picked using a few seed points and an auto-picker.

2d-merge,2d-topbase
Figure 2. (a) 2D segmentation result without interpreter guidance. The salt body segment shows significant ``leakage'' at the locations indicated. (b) Manual interpretations of the salt boundary near where the leakages occurred.

At this point, these manual interpretations must somehow be incorporated into the automated framework of the PRC method. Broadly speaking, this can be done in two ways. First, we can modify the algorithm itself to check for the manual interpretations when assigning weights to pixel pairs. However, this option adds unnecessary complexity to the process and would likely severely curtail the algorithm's efficiency (one of its major advantages). Alternatively, we could modify the input data according to the picks, and allow the segmentation algorithm to proceed normally. Since this option preserves the algorithm's efficient structure, I focus here on modifying the input amplitude of the envelope volumes to accommodate the manual interpretations.

2d-env-orig,2d-env-new
Figure 3. Amplitude of the envelope calculations from the original image (a), and modified according to the manual salt picks in Figure 2(b). The salt boundary is clearer and more continuous in the modified image, especially at the indicated locations.

A simple solution to the problem would involve assigning a high amplitude value to the manually picked points. While this is the basic idea behind the solution described here, it is unfortunately not quite so straightforward. Assigning high amplitudes to a string of points without regard to neighboring amplitude values causes the algorithm to treat the manual interpretation as a segment distinct from those surrounding it, even those also inside a salt body. Instead, I define a new amplitude value $(A)$ for a ``picked'' pixel at position (x,y,z) in terms of the highest-amplitude pixel in a neighborhood surrounding it and a scaling factor $\alpha$ :

$$A_{xyz} = \alpha\max_{ijk}A_{ijk},$$ (1)

where $i$ , $j$ , and $k$ are constrained by ${\vert x-i\vert\leq5, \vert y-j\vert\leq5, \vert z-k\vert\leq5}$ . This approach ensures that a boundary will be emphasized on the input data where it has been manually picked, but will not appear radically different from its surroundings. Figures 3(a) and 3(b) illustrate this idea; the manually picked points are now clearly visible in panel (b), but the character of the boundary does not change noticeably. Now, segmenting the new input image with parameters identical to the original segmentation yields the result seen in Figure 4. The segments conform to the manual picks seen in Figure 2(b), while the rest of the image is segmented as accurately as the original result in Figure 2(a).

2d-merge-new
Figure 4. 2D interpreter-guided segmentation results. This result incorporates the manual picks seen in Figure 2(b).

3d-origseg,oct-3d-picks
Figure 5. (a) 3D segmentation result prior to any interpreter guidance. (b) Manually interpreted salt boundary picks. Manual picks from one or more sections are used to guide the automatic 3D segmentation.

pyramid Figure 6. Schematic illustrating the concept of ``spraying" manual interpretations of a 2D section into the third dimension. The black dot represents a pixel along a manually-picked horizon.

3d-env-orig,3d-env-new
Figure 7. 3D amplitude of the envelope calculations from (a)the original image, and (b) modified according to the manual picks seen in Figure 5(b). The 3D effects of the procedure to incorporate the picks are evident on the crossline and depth slices.

3d-newseg
Figure 8. 3D interpreter-guided segmentation results, incorporating the picks seen in Figure 5(b). Note that in this result, the segmentation is more accurate in all three directions.

Interpreter input for seismic image segmentation