Banca de DEFESA: THIAGO THEIRY DE OLIVEIRA

Uma banca de DEFESA de MESTRADO foi cadastrada pelo programa.
STUDENT : THIAGO THEIRY DE OLIVEIRA
DATE: 06/07/2026
TIME: 09:00
LOCAL: Auditório do Núcleo de Pesquisa e Inovação em Tecnologia da Informação - nPITI
TITLE:

Automated mapping of the boundary between continental and oceanic crusts (COB/COT) based on Machine Learning.

KEY WORDS:

Computer Vision, Machine Learning, Seismic Interpretation, Neural Networks, Continental Margin.

PAGES: 80
BIG AREA: Engenharias
AREA: Engenharia Elétrica
SUMMARY:

The precise delimitation of the transition between continental crust and oceanic crust, known as the Continent-Ocean Transition (COT) and the Continent-Ocean Boundary (COB), represents a complex challenge for understanding the internal structure of the Earth and the geodynamic processes governing its evolution over geological time. The systematic mapping of the COB/COT has attracted increasing interest from the geoscientific community due to the structural complexity involved in the tectonic evolution of continental margins and the gradual nature of this lateral transition between different types of crust. The identification of this crustal interface in the subsurface provides essential information on tectonic evolution, sedimentary basin formation, and crustal architecture, as well as having relevant scientific, economic, and geopolitical implications, particularly in the context of natural resource exploration and the definition of territorial and maritime boundaries. Traditionally, the delimitation of the COB/COT requires a multidisciplinary approach based on the integration of different types of geological, geophysical, and geospatial data, and is strongly dependent on the subjective interpretation of geoscientists. This subjective nature makes it necessary to develop more objective, reproducible, and efficient methodologies. In this context, machine learning techniques emerge as a promising alternative for identifying complex patterns and subtle relationships in multidisciplinary data that indicate the presence of the continental–oceanic crust transition. This dissertation proposes the development of an automated methodology for COB/COT mapping based on seismic data, through the classification of seismic image blocks (patches) previously labeled by geologists into five geological classes: water, sedimentary basin, continental crust, oceanic crust, and mantle. Initially, an approach based exclusively on Convolutional Neural Networks (CNNs) is evaluated, followed by methodological extensions that explicitly incorporate the spatial position of the patches and Transformer based architectures, including hybrid models that combine CNNs with self-attention mechanisms. The dataset was constructed from seismic images interpreted by geologists, organized into blocks (patches) that enable the application of different machine learning strategies. The proposed methodologies allow the spatial reconstruction of the segmentation of the original seismic image from predictions performed on the patches, preserving the spatial coherence of the mapped geological units. The explicit incorporation of spatial information and attention mechanisms proved to be essential for enriching data representation, facilitating the identification of patterns associated with the continental–oceanic crust transition, especially in regions of greater geophysical complexity. Overall, the conducted experiments demonstrate that deep learning–based approaches, particularly hybrid CNN–Transformer architectures, constitute effective tools to support COB/COT delimitation. The obtained results indicate satisfactory performance, with an overall accuracy of approximately 92% on the test set, demonstrating good discrimination capability among the main geological units. It is observed that the combined use of convolutional operations and self-attention mechanisms contributes to improving classification robustness and reducing interpretative subjectivity, thereby expanding the potential application of artificial intelligence techniques in seismic analysis. Thus, the proposed methodology opens perspectives for future studies with larger datasets and different geological contexts.

COMMITTEE MEMBERS:
Presidente - 347628 - ADRIAO DUARTE DORIA NETO
Interno - 2579664 - ALLAN DE MEDEIROS MARTINS
Externo ao Programa - 1315614 - DAVID LOPES DE CASTRO - UFRNExterna à Instituição - ALANNY CHRISTINY COSTA DE MELO