Distinguishing Between Gas Chimneys and Shallow Gas-Bearing Sediments

This study addresses the critical issue of seismic data interpretation: accurate identification of gas chimneys in the presence of interfering shallow gas-bearing sediments. The conventional machine-learning methods, especially multilayer perceptron (MLP), have been popular for gas-chimney detection but poorly distinguish chimney signatures from those surrounded by features with similar characteristics. To overcome these limitations, we have applied a Convolutional Neural Network (CNN) method that directly learns spatial relationships from the data. We trained a CNN model on seismic sections using amplitude data combined with similarity attributes and signal-to-noise ratio as input features. In an effort to tackle the limited number of training examples, we followed a patch-based approach with 25×49 sample dimensions. The model, when tested on the North Sea F3 block dataset, resulted in 95% training accuracy and 93% validation accuracy. Results show significantly improved discrimination between true gas chimneys and non-chimney intervals, such as shallow gas-concentrated sediments and dewatering zones. This enhanced detection capability allows for better characterization of near-subsurface fluid migration pathways and associated drilling hazards, showing the potential of deep learning in advancing seismic interpretation tasks.