Assessment of texture descriptors for seismic image retrieval and salt dome detection

Abstract

Much work has been done on the assessment of texture descriptors for image retrieval many domains. In the context of geoscience, the image retrieval has been applied to automatically identify important structures in a seismic cube, like salt domes and fault. In this work, we evaluate the accuracy and performance of four well-known texture descriptors - namely, Gabor Filters, Grey Level Co-occurrence Matrix (GLCM), Local Binary Patterns (LBP), and Histogram Features (HF) - for seismic image retrieval and salt dome detection. These subsurface images pose challenges yet not thoroughly investigated in previous works, which are addressed and evaluated in our experiments. We asked for domain experts to annotate two seismic cubes - Penobscot 3D and Netherlands F3 - and used them to evaluate texture descriptors, corresponding parameters, and similarity metrics with the potential to retrieve similar regions and detect salt domes in the considered datasets. While GLCM is used in the vast majority of works in geosciences, our findings indicate that LBP has the potential to produce satisfying results for seismic image retrieval with lower computational cost. By the same token, HF had a good impact in salt-dome detection.