Soft Ensemble U-Net dan Vision Transformer dengan Integrasi Multi-Atribut Seismik untuk Horizon Picking Otomatis pada Dataset F3 Netherlands

Zakaria, Moch Lutfi (2026) Soft Ensemble U-Net dan Vision Transformer dengan Integrasi Multi-Atribut Seismik untuk Horizon Picking Otomatis pada Dataset F3 Netherlands. Masters thesis, Institut Teknologi Sepuluh Nopember.

Text 6025221048-Master_Thesis.pdf - Accepted Version Restricted to Repository staff only Download (4MB) | Request a copy

Abstract

Interpretasi seismik merupakan tahapan krusial dalam eksplorasi minyak dan gas bumi, di mana proses horizon menjadi fundamental untuk pemetaan reservoir. Metode horizon picking manual memiliki keterbatasan signifikan, yaitu durasi interpretasi yang panjang dan permasalahan skalabilitas pada volume data yang terus meningkat. Pendekatan deep learning berbasis arsitektur tunggal seperti CNN dan U-Net telah menunjukkan hasil yang menjanjikan, namun memiliki keterbatasan dalam menangkap dependensi spasial global serta belum optimal dalam mengintegrasikan informasi geologis yang komprehensif dari multiatribut seismik. Penelitian ini mengembangkan dan mengevaluasi arsitektur ensemble deep learning yang menggabungkan U-Net dan Vision Transformer (ViT) untuk horizon picking pada dataset F3 Netherlands, dengan integrasi multiatribut seismik (amplitude, instantaneous phase, envelope, coherence, dan dip). Melalui studi ablasi sepuluh konfigurasi atribut untuk masing-masing arsitektur, ditemukan bahwa performa U-Net mencapai mIoU 0,9007 dengan kombinasi empat atribut, sedangkan ViT justru optimal dengan satu atribut tunggal (mIoU 0,8118). Luaran terbaik dari kedua model kemudian digabungkan melalui teknik soft ensemble sehingga mampu menghasilkan mIoU 0,9055 pada bobot α=0,68, dimana performa ini lebih unggul daripada U-Net dan ViT. Kontribusi penelitian ini meliputi metodologi evaluasi untuk ensemble deep learning pada horizon picking, karakterisasi kebutuhan atribut seismik yang bergantung pada arsitektur, serta bukti empiris bahwa soft ensemble tanpa parameter terlatih dapat menjadi alternatif untuk menambah performa model.
================================================================================================================================
Seismic interpretation is a crucial stage in oil and gas exploration where the horizon process is fundamental for reservoir mapping. Manual methods face significant limitations, including long interpretation time, inconsistent results across interpreters, and limited scalability for potentially growing data volumes. Deep learning approaches based on single architectures such as CNN and U-Net have shown promising results, but are limited in capturing global spatial dependencies and are not yet optimal at integrating comprehensive geological information from multi-attribute seismic data. This research develops and evaluates an ensemble deep learning architecture that combines U-Net and Vision Transformer (ViT) for horizon picking on the F3 Netherlands dataset, integrating multi-attribute seismic data (amplitude, instantaneous phase, envelope, coherence, and dip). Through an ablation study of ten attribute configurations for each architecture, U-Net achieved an mIoU of 0.9007 with a four-attribute combination, whereas ViT was optimal with a single attribute (mIoU 0.8118). The best outputs of both models were then combined using a soft ensemble technique, yielding an mIoU of 0.9055 at a weight of α=0.68, a performance that outperforms both U-Net and ViT. Contributions of this research include an evaluation methodology for deep learning ensembles in horizon picking, a characterization of architecture-dependent seismic attribute requirements, and empirical evidence that an untrained soft ensemble can be an alternative to improving model performance.

Item Type:	Thesis (Masters)
Uncontrolled Keywords:	CNN Transformer Hybrid, Ensemble Deep Learning, Horizon Picking, Multi-Atribut Seismik, Dataset F3 Netherlands Ensemble Deep Learning, F3 Netherlands Dataset, Horizon Picking, Multi-Attribute Seismic, U-Net, Vision Transformer
Subjects:	T Technology > T Technology (General) > T58.5 Information technology. IT--Auditing
Divisions:	Faculty of Intelligent Electrical and Informatics Technology (ELECTICS) > Informatics Engineering > 55101-(S2) Master Thesis
Depositing User:	Moch Lutfi Zakaria
Date Deposited:	01 Aug 2026 03:05
Last Modified:	01 Aug 2026 03:05
URI:	http://repository.its.ac.id/id/eprint/141048

Actions (login required)

View Item