Banca de QUALIFICAÇÃO: MARCOS VINICIUS GOMES JACINTO

Uma banca de QUALIFICAÇÃO de MESTRADO foi cadastrada pelo programa.
STUDENT : MARCOS VINICIUS GOMES JACINTO
DATE: 05/05/2021
TIME: 13:30
LOCAL: videoconferencia - https://meet.google.com/usg-tjeo-cgf
TITLE:

KARSTIFIED FEATURES INTERPRETATION THROUGH MACHINE  LEARNING ALGORITHMS: CONVOLUTIONAL NEURAL NETWORKS APPLICATIONS TO  GROUND PENETRATING RADAR IMAGES


KEY WORDS:

Deep Learning; Convolutional Neural Networks; Ground Penetrating Radar


PAGES: 50
BIG AREA: Ciências Exatas e da Terra
AREA: Geociências
SUMMARY:

Most of recent papers that study Ground Penetrating Radar (GPR) and Machine Learning (ML)  applications are applied to civil engineering, lacking studies applied to geosciences. In this  context, the present work seeks to help fill this gap by studying automatic GPR interpretation  with machine learning methods in the context of a karstified environment with carbonate rocks  from the Irece Basin (Brazil). In the case of this work, there are two classes that represent two  different expected interpretations. The first one represents a karstified feature in carbonates,  while the second is anything but the karstified feature. The present work had access to eight  GPR sections. From these sections, the following attributes were generated: Energy,  Similarity, Instantaneous Amplitude, Instantaneous Phase, Instantaneous Frequency, Hilbert  Trace, Maximum Spectral Amplitude, Quality Factor, Hilbert Trace/Energy, and Hilbert  Trace/Similarity. To achieve this automatic interpretation, a state-of-art Deep Learning  algorithm, named U-Net Convolutional Neural Network, commonly applied to images is used,  combined with a feature selection method through Genetic Algorithms used to select the best  subset of features that will impact positively the model’s performance. The obtained results  show that it is possible, in fact, to generate automatic models for interpreting GPR sections  using U-Net. Furthermore, by continually comparing the results using different sets of  solutions, it was found that the methodology and model are robust enough and allow more  than one possible solution to obtain an evaluation metric above 95% when interpreting  karstified regions


BANKING MEMBERS:
Presidente - 350640 - FRANCISCO HILARIO REGO BEZERRA
Interno - 1315614 - DAVID LOPES DE CASTRO
Externo ao Programa - 347628 - ADRIAO DUARTE DORIA NETO
Notícia cadastrada em: 04/05/2021 13:59
SIGAA | Superintendência de Tecnologia da Informação - (84) 3342 2210 | Copyright © 2006-2024 - UFRN - sigaa09-producao.info.ufrn.br.sigaa09-producao