Lateral discontinuity of a reflector causes a lot of auto-picking algorithms to fail. Most auto-picking algorithms work based on comparisons of local features and ignore global trends. We have introduced a new algorithm in GVERSE Geophysics 2017.3 to better handle the unpredictability associated with picking across lateral discontinuities.

Consider a case where the event we are tracking encounters a fault with a throw equal to some multiples of seismic frequency. If we base our decision on that small portion of data, we will, most likely, not be able to predict where the event connects. But if we look at a larger portion of data we get a more complete view of these displaced events.

The R&D team at LMKR has been experimenting with an automatic 2D horizon tracking algorithm that can better handle lateral discontinuity. Our experiments have resulted in the development of our latest algorithm: the Segment Auto Pick. Our 2017.3 release has a new auto-picking algorithm that can better handle the unpredictability associated with picking across lateral discontinuities.

This new method of horizon tracking uses a global optimization-based technique as opposed to methods of tracking using local optimization-based techniques. In local techniques, a sample is picked based on similarity with its neighboring trace or traces. Because localized tracking relies only on the picked trace and its neighbors, a wrong pick will propagate the error to all other samples from the erroneous point onward. In our new auto-picker, we track the optimal horizon in a seismic section by maximizing the overall (sum of similarities) between seed and the tracked samples – what we call the cost function.

Our approach is to first explore all possible paths between the start and end points. The start and end points are input by the interpreter. All possible paths between the start and end points are computed and the algorithm then selects the path with the maximum overall cost function. Picking the highest cost path means selecting the samples that together, not locally, but globally have the most similarity to the seed points.

The algorithm works based on the following rules:

- Take two seed points from the interpreter to guide the process
- Fit a straight horizon between two seed points by linear interpolating the two seed points
- Fit a window of specific size having the height specified by user and length equal to the number of traces between the end points, centered and aligned along the straight line horizon computed in step 2.
- The seismic section inside the window is seen as graph considering each pixel/point in that section as a node.
- One of the picked point is used as start point and the other point as final point to stop the tracker and weighted average of the two traces is used as a reference trace to guide similarity measures
- The cost of each graph edge is given by correlation value:

** ****Cost(i,j) =**** corrScore(j) (1)**

Where:

**cost(i; j)**is the cost of the directed edge from node i to node j,**corr(j)**is the correlation coefficient between node j and the reference traces ( weighted average trace of picked points).

Based on these rules we start building a graph whose final state can be like the figure below

As shown in the figure each node of the graph keeps the best accumulated cost and knows the sample points of the best path from the origin seed point to itself. We start building the graph from the start point directed towards the end point. Each node generates three children in the next trace. At the start point we have only one a single pick and it picks three children in neighboring trace. The local costs and accumulated costs are calculated using equation (1) for each of these picks. Now each of these three child nodes can have further three children in the next trace and so on.Some of these children will always be overlapping i.e. nodes that can be tracked from multiple paths. The multiple paths are eliminated using the optimal path. The optimal path is determined based on parents that have the maximum accumulated cost. In this way we make sure that every node has been tracked by optimal path from the seed point.

After completing this graph, selecting the path connecting the seed point with the control point specified by the interpreter gives us the optimal path – as shown by green colored path in figures 2 and 3.

Outputs showed satisfactory results to correctly map horizons even when tracking among complicated seismic features like faults, discontinuities and at the vicinity of salt domes as shown in figures 2 and 3.

Although the tracker does not skip the fault in the correct positions on horizons in Figure 1 and Figure 3, this error is not propagated because the algorithm looks for the maximum similarity throughout the tracked horizon.

It is important to mention that since correlation is always computed with the seed samples, the quality of the results is directly related to the quality of the seed. Seed points close to the horizons will always give better results than points far away from the desired horizon.

The new auto-picker is part of the GVERSE Geophysics 2017.3 along with many other exciting features and improvements. Make sure you check out our videos and other blogs about all the new functionality that is part of this release.

References: **2D Horizon Tracking Using Dynamic Programming **Eliana Goldner, Pedro Mario Silva, Marcelo Gattass

## Leave a Reply