The role of visual saliency in the automation of seismic interpretation

 Collaborators: Amir Shafiq, Tariq AlShawi, Zhiling Long, Ghassan AlRegib

Goal/Motivation: To incorporate human visual system modeling in devising a method to perform seismic interpretation

Challenges: Various computational algorithms for seismic interpretation have been developed over the years. They focus on either detection or tracking of subsurface structures such as faults, salt domes, gas chimneys, and channels. They typically rely on analyzing characteristics observed in the imaging data, such as correlations, changes in intensity or contrast, and texture patterns, to name a few. Since seismic imaging data is represented in the form of 2-dimensional sections or 3-dimensional volumes, and visual inspection is utilized in a traditional interpretation by human interpreters, image or video processing algorithms have been heavily adopted in computational methods designed to assist seismic interpretation. However, saliency detection based on human visual attention, although being studied extensively in the past two decades in the image/video processing community, has rarely been applied to computational seismic interpretation.

High Level Description of the Work: Saliency detection techniques aim to identify visually prominent areas or objects within images and videos, which are salient areas or objects that most likely draw the attention of the human visual system (HVS). When applied to an image (or a video), a saliency detection algorithm will calculate a saliency value for each pixel (or voxel), resulting in a saliency map, which maps each pixel (or voxel) in the image (or video) to its associated saliency value. Correspondingly, areas in the image (or video) with high saliency values are identified as salient areas. Saliency detection mimics the attention mechanism of HVS, which enables HVS to focus its processing resources on elements conveying the most important information in a complex surrounding environment. Such a mechanism is described computationally by an attention model. 

Given that seismic interpretation is traditionally accomplished by human interpreters through visual inspection, saliency detection algorithms, if designed appropriately to imitate a human interpreter’s visual attention, should be able to identify in seismic imaging data geological structures of interest to an interpreter as salient areas. Such algorithms will provide a brand new perspective to examine seismic imaging data. 

We present a novel saliency detection algorithm for seismic imaging data in the form of 3D data volumes, which incorporates domain knowledge specific for seismic interpretation into the underlying attention model. Following our previous study reported in [1, 2], we establish our new algorithm on the basis of the same 3D-FFT-based saliency detection method. The utilization of 3D-FFT helps capture energy variations within a seismic volume both effectively and computationally efficiently, and the multi-dimensional spectral projection tracks the variations along different dimensions reliably. Our major innovation in this new saliency detection algorithm is with the underlying attention model, in which we propose a unique directional center-surround comparison to replace the non-directional comparison commonly performed for natural images and videos. This directional operation is adjustable for any orientation, and aims to better accommodate seismic structures of directional nature, such as faults, gas chimneys, and horizons. Additionally, we design an adaptive scheme to automatically combine saliency values calculated from the multi-dimensional spectral projections into an overall saliency value. For evaluation, we use two entirely different seismic datasets under different geological settings to examine the effectiveness of the proposed saliency detection algorithm. We demonstrate that our innovative method outperforms comparable state-of-the-art algorithms in detecting salient subsurface structures of interest in seismic interpretation. 

References: 

  1. Shafiq, Muhammad Amir, et al. "Salsi: A new seismic attribute for salt dome detection." 2016 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2016. [PDF][Code]

  2. Shafiq, Muhammad Amir, et al. "The role of visual saliency in the automation of seismic interpretation." Geophysical Prospecting 66.S1 (2018): 132-143. [PDF][Code]