Segmentação de acumulações de gás natural em imagens de reflexão sísmica através de técnicas 1D, 1.5D e 2D

Abstract

The most common indirect identification method used by the oil and natural gas industry in the search for natural gas accumulations is conducting a survey based on the principle of seismic reflection. Specialists look for sudden changes in signal intensity, which may indicate the presence of natural gas, oil, or other hydrocarbons. In this thesis, we propose three methods to automate this task using seismic reflection data: one method that uses an Ensemble of U-Nets, a well-established architecture in image segmentation, and two others that make use of the Robustly Optimized Bidirectional Encoder Representation from Transformers Pretraining Approach (RoBERTa), a method originally developed in the context of natural language modeling. The methods are developed in such a way that we simulate what happens in practice, where models are trained in a given region and are used to evaluate unkown areas. To do so, we propose a train-test split policy for the public dataset used that simulates the use of the methods in a new region. The U-Net based method requires the annotation to be considerably larger than the original and is able to determine the annotated region more precisely. The method based on the RoBERTa architecture shows more false positives in areas of high acoustic-impedance

variation, but it is capable of identifying possible smaller accumulations, as well as de- lineating much of the structure that stores the natural gas. All three methods present

segmentation evaluation metrics superior to the state-of-the-art when considering the model’s generalization capability using a single dataset. On the F3 block dataset in the Netherlands, the methods achieved F1-scores between 0.36 and 0.62 on the proposed test set, while sensitivity varies between 0.65 and 0.79.