Perancangan Sistem Deteksi Dini Gempa Bumi Berbasis Residual Neural Network untuk Estimasi Magnitudo Gempa

Leksana, Jordan Arya (2026) Perancangan Sistem Deteksi Dini Gempa Bumi Berbasis Residual Neural Network untuk Estimasi Magnitudo Gempa. Other thesis, Institut Teknologi Sepuluh Nopember.

Text 5009221116-Undergraduate_Thesis.pdf - Accepted Version Restricted to Repository staff only Download (8MB) | Request a copy

Abstract

Indonesia merupakan wilayah dengan aktivitas seismik tinggi sehingga pengembangan sistem deteksi dini gempa bumi menjadi aspek penting dalam upaya mitigasi bencana. Penelitian ini mengembangkan sistem deteksi prekursor gempa bumi dan estimasi kelas magnitudo berbasis sinyal geomagnetik Ultra Low Frequency (ULF) pada pita Pc3 menggunakan pendekatan deep learning. Data geomagnetik komponen H, D, dan Z dari jaringan stasiun BMKG tahun 2018–2025 diproses melalui tahapan seleksi data, penyaringan sinyal, dan transformasi Continuous Wavelet Transform (CWT) untuk menghasilkan representasi scalogram. Model yang dikembangkan menggunakan arsitektur Residual Neural Network (ResNet-18) dengan pendekatan multi-task learning multi-stasiun, sehingga mampu memanfaatkan informasi dari beberapa stasiun pengamatan untuk menghasilkan prediksi pada tingkat kejadian. Model terdiri atas dua head keluaran, yaitu head deteksi prekursor dan head klasifikasi magnitudo. Hasil pengujian menunjukkan bahwa model mampu mendeteksi keberadaan prekursor gempa dengan akurasi sebesar 88,25% serta nilai precision dan recall yang relatif seimbang pada kedua kelas yang ditandai dengan F1-Score sebesar 88,4%. Interpretasi menggunakan Gradient-weighted Class Activation Mapping (Grad-CAM) menunjukkan bahwa model memiliki pola aktivasi yang stabil dan konsisten antar-stasiun pada tugas deteksi prekursor. Sementara itu, estimasi magnitudo tiga kelas, yaitu Moderate, Medium, dan Large, memperoleh akurasi sebesar 52,67% dan masih menunjukkan keterbatasan akibat ketidakseimbangan distribusi kelas serta lemahnya hubungan langsung antara amplitudo anomali geomagnetik dan magnitudo gempa. Penelitian ini menunjukkan bahwa pemanfaatan deep learning pada analisis sinyal geomagnetik berpotensi mendukung pengembangan sistem deteksi dini gempa bumi. Selain itu, pendekatan yang dikembangkan sejalan dengan SDGs poin 9 melalui penerapan inovasi berbasis kecerdasan buatan, serta SDGs poin 13 dalam upaya meningkatkan ketangguhan masyarakat terhadap risiko bencana.
======================================================================================================================================
Indonesia is located in a region with high seismic activity; therefore, the development of an earthquake early detection system is an important aspect of disaster mitigation. This study develops an earthquake precursor detection and magnitude class estimation system based on Ultra Low Frequency (ULF) geomagnetic signals in the Pc3 band using a deep learning approach. Geomagnetic data consisting of H, D, and Z components from the BMKG station network from 2018 to 2025 were processed through data selection, signal filtering, and Continuous Wavelet Transform (CWT) to generate scalogram representations. The proposed model uses a Residual Neural Network (ResNet-18) architecture with a multi-station multi-task learning approach, enabling the model to utilize information from several observation stations and produce event-level predictions. The model consists of two output heads, namely the precursor detection head and the magnitude classification head. The testing results show that the model is able to detect earthquake precursors with an accuracy of 88.25%, with relatively balanced precision and recall values for both classes showed with a F1-Score of 88,4%. Model interpretation using Gradient-weighted Class Activation Mapping (Grad-CAM) indicates stable and consistent activation patterns across stations in the precursor detection task. Meanwhile, the three-class magnitude estimation task, consisting of Moderate, Medium, and Large classes, achieved an accuracy of 52.67% and still showed performance limitations due to class distribution imbalance and the weak direct relationship between geomagnetic anomaly amplitude and earthquake magnitude. These results indicate that deep learning-based analysis of geomagnetic signals has potential to support the development of earthquake early detection systems. Furthermore, the proposed approach aligns with SDG 9 through the application of artificial intelligence-based innovation and SDG 13 by supporting efforts to improve community resilience against disaster risks.

Item Type:	Thesis (Other)
Uncontrolled Keywords:	Continuous Wavelet Transform, Deteksi Dini Gempa, Prekursor ULF, Residual Neural Network, Sinyal Geomagnetik,Continuous Wavelet Transform, Earthquake Early Detection, Geomagnetic Signal, Residual Neural Network, ULF Precursor
Subjects:	Q Science Q Science > QE Geology > QE1.F557 Magnetic anomalies--Measurement. Q Science > QE Geology > QE538.8 Earthquakes. Seismology T Technology > T Technology (General) > T58.5 Information technology. IT--Auditing
Divisions:	Faculty of Industrial Technology and Systems Engineering (INDSYS) > Physics Engineering > 30201-(S1) Undergraduate Thesis
Depositing User:	Jordan Arya Leksana
Date Deposited:	04 Aug 2026 04:06
Last Modified:	04 Aug 2026 04:06
URI:	http://repository.its.ac.id/id/eprint/140662

Actions (login required)

View Item