Automatic gas chimney detection from 3D seismic reflection data using a single amplitude attribute

Abstract

Automatic gas chimney classification methods use selected seismic attributes as training inputs for machine-learning algorithms. Contrary to the common belief that a single attribute is insufficient, the scope of this paper is to introduce an algorithm for gas chimney detection and extraction based on the amplitude attribute. The method is applied to a 3D seismic cube and involves the use of a feature extraction process (maximum overlap discrete wavelet transform - MODWT), before applying a standard machine-learning algorithm (bidirectional long short-term memory - BiLSTM). The MODWT is used to extract four unique features of each trace, which are then used with the original trace for training. The idea is to introduce a new feature-extraction step that enhances the accuracy of a single attribute. When tested on the F3 Block seismic cube, the integration of this process with the BiLSTM deep-learning algorithm achieved a balanced accuracy of 93% in gas chimneys identification. Results calibration by an experienced seismic interpreter also show a good match between machine-learning algorithm and manual interpretation. Therefore, our results successfully demonstrate that gas chimneys can be automatically detected and extracted from 3D seismic data using a single attribute. The workflow presented here is applicable worldwide for studies of different economic and environmental purposes (e.g., natural resources of petroleum exploration and gas hydrate, subsurface storage characterization and geohazards identification).