Fahriza, Achmad (2026) Pemodelan Inundasi Tsunami 3D Resolusi Tinggi Berbasis Rekonstruksi Kawasan Permukiman Urban dan Wisata Pesisir dari Multi-Source Stereo Imagery Menggunakan Deep Neural Network. Other thesis, Institut Teknologi Sepuluh Nopember.
![[thumbnail of 5016221008-Undergraduate_Thesis.pdf]](http://repository.its.ac.id/style/images/fileicons/text.png) |
Text
5016221008-Undergraduate_Thesis.pdf - Accepted Version Restricted to Repository staff only Download (15MB) | Request a copy |
Abstract
Zona megathrust selatan Jawa yang menyimpan defisit slip lebih dari 300 tahun menempatkan pesisir selatan Bali dalam ancaman tsunami katastrofik dengan jendela evakuasi sesingkat 9-23 menit, namun seluruh model inundasi yang ada masih merepresentasikan bangunan sebagai nilai kekasaran ekuivalen alih-alih hambatan geometrik eksplisit, suatu penyederhanaan yang terbukti menghasilkan bias flow depth hingga 35-60% pada kawasan urban padat. Penelitian ini mengembangkan pipeline rekonstruksi bangunan LoD2 dari 15 scene citra PlanetScope L1B cross-date (GSD 3,81-4,75 m) berbasis deep neural network, membakar model bangunan yang dihasilkan ke dalam domain inundasi resolusi 1 meter, lalu mensimulasikan enam skenario gempa megathrust Mw 7,8-8,9 menggunakan COMCOT 1.7. Pipeline mengintegrasikan RAFT-Stereo (61 pasangan valid, CA = 4,00°, B/H = 0,0698), ensemble TabularMLP-LightGBM-NNLS (135 fitur, bobot MLP:LGB = 0,888:0,112), kalibrasi isotonik berbasis 275 bangunan beranchor LiDAR, dan CNN EfficientNet-B0 untuk klasifikasi tipe atap. Dari 714.642 bangunan dalam domain ±284 km², metode menghasilkan RMSE ketinggian out-of-sample 0,703 m (CE90 = 0,761 m; R² = 0,919), melampaui benchmark SOTA 1,5-3,0 m dengan faktor lebih dari dua, serta F1-Score segmentasi 0,866 yang melampaui GBA, Google OB, VIDA, dan Microsoft. Validasi terhadap tiga kejadian historis (Sumba 1977 MAPE = 31,54%; Banyuwangi 1994 MAPE = 28,07%; Pangandaran 2006 MAPE = 23,03%) dengan seluruh 27 titik dalam batas faktor dua mengkonfirmasi keandalan model. Skenario Mega-Sumba (Mw 8,5) terbukti paling kritis meskipun bukan terbesar magnitudonya, dengan ETA minimum 9,47 menit di Pandawa, run-up proyeksi 17,026 m di Tanjung Benoa, dan status Awas di seluruh sepuluh stasiun, membuktikan bahwa kedekatan sumber dan fokus batimetri lebih menentukan distribusi bahaya lokal daripada magnitudo. Asesmen dampak berbasis fragility curve memproyeksikan kerugian gabungan USD 5,97-14,74 miliar (41-101% PDRB Bali 2024) dengan kerugian pariwisata mendominasi pada rasio 2,4:1. Penelitian ini membuktikan kelayakan rekonstruksi kawasan permukiman pesisir dari citra satelit stereo sebagai alternatif LiDAR airborne sekaligus menyediakan landasan bagi mitigasi tsunami pesisir selatan Bali.
=====================================================================================================================================
The southern Java megathrust, carrying over three centuries of unrelieved slip deficit,threatens the southern coast of Bali with catastrophic tsunami inundation and evacuationwindows as short as 9-23 minutes. Yet every existing model still parameterises buildings asequivalent roughness values rather than explicit flow obstacles, a simplification proven tointroduce flow-depth biases of 35-60% in dense urban areas. This study develops a deep neuralnetwork LoD2 building reconstruction pipeline from 15 cross-date PlanetScope L1B stereoscenes (GSD 3.81-4.75 m), burns the resulting models into 1-meter inundation terrain, andsimulates six megathrust scenarios (Mw 7.8-8.9) using COMCOT 1.7. The pipeline combinesRAFT-Stereo dense matching (61 valid pairs, CA = 4.00°, B/H = 0.0698), a 135-featureTabularMLP-LightGBM-NNLS stacking ensemble (weights 0.888:0.112), LiDAR-anchored 5-fold isotonic calibration, and EfficientNet-B0 roof classification. Across 714,642 buildings in±284 km², the method achieves an out-of-sample height RMSE of 0.703 m (CE90 = 0.761 m;R² = 0.919), exceeding the state-of-the-art benchmark of 1.5-3.0 m by more than a factor oftwo, and a segmentation F1-Score of 0.866 surpassing GBA, Google OB, VIDA, and Microsoft.Validation against three historical events Sumba 1977 (MAPE = 31.54%), Banyuwangi 1994(MAPE = 28.07%), and Pangandaran 2006 (MAPE = 23.03%) with all 27 points within factor-of-two bounds confirms model reliability. The Mega-Sumba scenario (Mw 8.5) proves mostcritical despite its lower magnitude, yielding a minimum estimated time of arrival of 9.47minutes at Pandawa, projected run-up of 17.026 m at Tanjung Benoa, and Warning status at allten gauges, demonstrating that source proximity and bathymetric focusing dominate localhazard distribution over magnitude alone. Fragility-curve-based impact assessment projectscombined losses of USD 5.97-14.74 billion (41-101% of Bali's 2024 GRDP), with tourismrevenue losses dominating at a 2.4:1 ratio. This study establishes medium-resolution satelliteLoD2 reconstruction as a viable alternative to airborne LiDAR and provides the mostcomprehensive quantitative tsunami hazard framework yet produced for southern Bali.
| Item Type: | Thesis (Other) |
|---|---|
| Uncontrolled Keywords: | Rekonstruksi Bangunan LoD2, Bahaya Tsunami Megathrust, Fragility Curve Tsunami, Inundasi Explicitly Represented Building, Stereo Satellite Deep Learning, LoD2 Building Reconstruction, Megathrust Tsunami Hazard, Tsunami Fragility Curve, Explicitly Represented Building, Stereo Satellite Deep Learning |
| Subjects: | Q Science > QA Mathematics > QA76.87 Neural networks (Computer Science) Q Science > QA Mathematics > QA911 Fluid dynamics. Hydrodynamics Q Science > QE Geology Q Science > QE Geology > QE538.8 Earthquakes. Seismology |
| Divisions: | Faculty of Civil, Planning, and Geo Engineering (CIVPLAN) > Geomatics Engineering > 29202-(S1) Undergraduate Thesis |
| Depositing User: | Achmad Fahriza |
| Date Deposited: | 28 Jul 2026 02:01 |
| Last Modified: | 28 Jul 2026 02:01 |
| URI: | http://repository.its.ac.id/id/eprint/138011 |
Actions (login required)
 |
View Item |