03 Aug
10:00 noon CEST Master's Defense Fully distance
Generative Adversary Networks for Noise Removal in Seismic Data
Jonlenes Silva de Castro
Advisor / Teacher
Sandra Eliza Fontes from Avila
Brief summary
Seismic data provide information on the characteristics of the geological environment, which, after being processed, are used in the interpretation to identify and locate the characteristics of the subsoil. The level of complexity of this task is directly influenced by the presence of noise, artifacts and / or inaccuracies in seismic data. Therefore, removing noise and increasing the quality of this data are essential steps in seismic processing. To mitigate these problems, we propose approaches based on generative adversarial networks (GANs) to improve seismic data by removing noise and increasing high frequencies in the data. In this context, we mapped the problem of noise removal as a problem of translation from image to image. In other words, we use GANs to learn how to map seismic data with noise into data without noise, maintaining the integrity of the data. However, training deep neural networks (such as GANs) has an inherent problem of scarcity of labeled data. To solve this problem, we developed a labeled synthetic seismic data generator. With this data, we created our models based on the architectures of Pix2Pix, Pix2PixHD and SPADE and used these models to perform the prediction of synthetic seismic data and real seismic data. Our results, both in synthetic and real seismic data, suggest that the proposed approaches are promising, showing significant improvements in the signal-to-noise ratio and in the gain of high frequencies. Thus, it is possible to use GANs as support tools for seismic specialists, in order to assist seismic analysis and decision making in a faster and more assertive manner.
Examination Board
Sandra Eliza Fontes from Avila IC / UNICAMP
Lúcio Tunes dos Santos IMECC / UNICAMP
Helio Pedrini IC / UNICAMP
Gerberth Adín Ramírez Rivera IC / UNICAMP
Jorge Henrique Faccipieri Junior CEPETRO / UNICAMP