Few-shot Learning for Semantic Segmentation of Seismic Data

Seismic interpretation i.e. the process of identifying objects of interest in the subsurface using seismic data, can be effectively represented as a semantic segmentation task in machine learning.

Contrary to many semantic segmentation tasks, seismic data present a major challenge with the amount of quality labeled data available. This stems from two factors: first, seismic data can be labeled almost exclusively by trained experts, and second, due to the subsurface’s heterogeneity, labels created for one location are not always transferable to another dataset.

We embrace this data shortage and reformulate the seismic segmentation problem from a fully supervised approach to a few-shot learning task. Our approach builds upon an existing few-shot learning method, which we adapt to the specific requirements of a multi-class subsurface segmentation problem. Furthermore, we address the image-patching locality problem by injecting a global view of the labels into the network during training and inference. Finally, we provide a computationally efficient post-processing approach that can be used with other existing seismic segmentation methods.

The proposed few-shot learning approach for semantic seismic segmentation outperforms a supervised UNet baseline implementation in terms of qualitative and quantitative results while being trained on a total of 10 inlines of each survey