Fault-Seg-LNet: A method for seismic fault identification based on lightweight and dynamic scalable network

Abstract

Improving the precision and efficiency of fault interpretation is of great significance for assisting in the search for oil-gas reservoir. At present, seismic image segmentation places more emphasis on the pursuit of model precision, while ignoring severe memory consumption and longer latency. Therefore, this paper proposes the seismic fault identification method based on lightweight and dynamic scalable network to assist in fault interpretation for actual geophysical prospecting work. Specifically, the fault identification model Fault-Seg-LNet based on multi-paths learning framework is proposed to connect convolutional streams in parallel from high-resolution to low-resolution, which can explicitly maintain high-resolution feature representation instead of restoring resolution from low-scale feature mapping. The multi-paths with multi-scales interaction (MPMSI) module is proposed to achieve repeated multi-scales feature interaction, which can strengthen the expression ability of the network in capturing fine-grained feature and global feature. The dynamic scalable paradigm is proposed to dynamically adjust the input resolution, network width and depth, which can achieve superior model segmentation precision while maintaining low complexity and high efficiency. Experimental results on synthetic dataset and field dataset show that the seismic fault identification method based on lightweight and dynamic scalable network can achieve the tradeoff between model precision and efficiency, and can locate and segment fault with different scales on seismic images.