Files
Abstract
This thesis develops deep learning-based semantic segmentation methods for seismic facies analysis and interprets the segmentation results to gain physical insights from the data.Many existing solutions incorporate image processing and machine learning to perform the segmentation required for domain applications.
These solutions require preprocessing steps, such as feature extraction and splitting images into tiles, to fit machine learning models and reduce the hardware costs caused by the high-resolution seismic data.
This thesis systematically analyzes the existing challenges and presents a novel end-to-end model to address the issues.
Seismic data is naturally noisy and frequently fails to provide optimal insights for interpretation.We decompose the data using a continuous wavelet transform to produce frequency-dependent spectral components and provide differently expressed geologic features at different frequencies. However, spectral components lead to two challenges: 1) the requirement for increased computing resources and 2) the leveraging of spectral features. Therefore, we propose a deep encoder-decoder-based network to alleviate those two challenges. In deep models, the feature extraction step is conducted by convolution, although the generated features lack suitable methods of interpretation and visualization. We employ a dilated convolution to reduce the demand on computing resources caused by the large dimensionality of spectral components. An attention mechanism is integrated into the deep encoder-decoder model to weigh the input spectral components, both spatial-wise and frequency-wise, thus overcoming the challenge of interpretation. The attention mechanism allows a neuron to selectively attend to the input data from its receptive field. This mechanism assists the model in focusing on a reduced set of features and suppresses counterparts. The resulting attention coefficients can be visualized as a salient map, revealing the impact of input contributions on the model outputs.
This end-to-end design of semantic segmentation of spectral components demonstrates how the latest computer vision and machine learning techniques can aid the solution of geoscience problems. The novel utilization of an attention mechanism demonstrates an alternative method of tracking and interpreting a deep model.
Seismic data is naturally noisy and frequently fails to provide optimal insights for interpretation.We decompose the data using a continuous wavelet transform to produce frequency-dependent spectral components and provide differently expressed geologic features at different frequencies. However, spectral components lead to two challenges: 1) the requirement for increased computing resources and 2) the leveraging of spectral features. Therefore, we propose a deep encoder-decoder-based network to alleviate those two challenges. In deep models, the feature extraction step is conducted by convolution, although the generated features lack suitable methods of interpretation and visualization. We employ a dilated convolution to reduce the demand on computing resources caused by the large dimensionality of spectral components. An attention mechanism is integrated into the deep encoder-decoder model to weigh the input spectral components, both spatial-wise and frequency-wise, thus overcoming the challenge of interpretation. The attention mechanism allows a neuron to selectively attend to the input data from its receptive field. This mechanism assists the model in focusing on a reduced set of features and suppresses counterparts. The resulting attention coefficients can be visualized as a salient map, revealing the impact of input contributions on the model outputs.
This end-to-end design of semantic segmentation of spectral components demonstrates how the latest computer vision and machine learning techniques can aid the solution of geoscience problems. The novel utilization of an attention mechanism demonstrates an alternative method of tracking and interpreting a deep model.