FaultSeg3D plus: A comprehensive study on evaluating and improving CNN-based seismic fault segmentation

• DOI: 10.1190/geo2022-0778.1
• Publication Date: 2024-05-29T15:08:14Z
• Authors: You Li; Xinming Wu; Zhenyu Zhu; Jicai Ding; Qingzhen Wang
• Keywords: 01 natural sciences; 0105 earth and related environmental sciences

Convolutional neural networks (CNNs) have been widely used for seismic fault segmentation and show more powerful performance than conventional attribute-based methods to obtain a map with noise-free continuously trackable features. However, CNN-based face the potential problem of poor generalization in field images, factors affecting remain incompletely studied or unexplored. Moreover, existing pixel-wise metrics, borrowed from natural image tasks, cannot fairly reasonably evaluate results. We first develop distance-based metric provide geologically reasonable evaluation on interpretation. then use most commonly U-net architecture as an example study how is affected by some significant such training data, many kinds network hyperparameters, scaling rotation inference step. Experimental results that data set realistic reflection features multiple sampling rates can enrich variations structure waveform signatures, thus significantly enhancing segmentation. novel loss function we developed outperforms others notable margins. Last, but not least, it necessary apply test-time augmentation strategy merging predictions scales rotations reference step because CNN does preserve transformation invariance. Based studies, optimally train properly designed examples, where accurate, clean, continuous detections, quantitatively them manual interpretations.