Segmentasi Semantik Kerusakan Bangunan Pascabencana Menggunakan Swin Transformer dengan Transfer Learning

Farid, Syifa Hayyina (2026) Segmentasi Semantik Kerusakan Bangunan Pascabencana Menggunakan Swin Transformer dengan Transfer Learning. Other thesis, Institut Teknologi Sepuluh Nopember.

Text 5002221017_Undergraduate_Thesis.pdf - Accepted Version Restricted to Repository staff only Download (37MB) | Request a copy

Abstract

Bencana alam dapat menyebabkan kerusakan bangunan yang berdampak langsung terhadap keselamatan masyarakat. Oleh karena itu, dibutuhkan metode identifikasi tingkat kerusakan bangunan yang cepat, akurat, dan mampu memberikan informasi spasial secara detail untuk membantu proses penanggulangan pascabencana. Penelitian ini mengimplementasikan Swin Transformer dengan decoder UPerNet untuk segmentasi semantik tingkat kerusakan bangunan pada citra satelit pascabencana menggunakan metode transfer learning dengan teknik fine tuning guna meningkatkan adaptasi model terhadap perbedaan karakteristik antarbencana. Tingkat kerusakan dipetakan ke dalam empat kelas, yaitu Tidak Rusak, Rusak Ringan, Rusak Sedang, dan Rusak Berat. Swin Transformer digunakan sebagai encoder karena mampu mengekstraksi fitur hierarkis lokal dan global secara efisien melalui mekanisme shifted window attention, sedangkan UPerNet digunakan sebagai decoder untuk menggabungkan fitur multi-skala sehingga menghasilkan peta segmentasi per piksel yang lebih detail. Untuk mendukung implementasi tersebut, penelitian ini menggunakan citra satelit pascabencana dari enam kejadian bencana dengan tahapan yang meliputi pengumpulan data, prapemrosesan citra, pelatihan model dasar pada domain sumber, dan fine-tuning pada domain target. Pelatihan pada domain sumber menghasilkan model dasar dengan mIoU sebesar 0.4537 dan F1-Score sebesar 0.6185. Model dasar tersebut kemudian digunakan sebagai bobot awal pada proses fine-tuning di domain target melalui beberapa skenario jumlah data, yaitu zero shot, 25% data, 50% data, dan 100% data untuk menilai pengaruh kuantitas data domain target terhadap performa model. Evaluasi performa model dilakukan menggunakan metrik mIoU, F1-Score, dan accuracy. Hasil penelitian menunjukkan bahwa metode transfer learning mampu meningkatkan kemampuan adaptasi model terhadap perbedaan karakteristik antarbencana. Pada skenario zero shot, model memperoleh mIoU sebesar 0.3861 dan F1-Score sebesar 0.5270. Setelah dilakukan fine-tuning, performa meningkat secara konsisten hingga mencapai mIoU sebesar 0.4831 dan F1-Score sebesar 0.6349 pada skenario 100% data domain target.
==================================================================================================================================
Natural disasters can cause building damage that directly impacts public safety. Therefore, a method for identifying building damage levels that is fast, accurate, and capable of providing detailed spatial information is needed to support post-disaster response efforts. This study implements Swin Transformer with decoder UPerNet for semantic segmentation of building damage levels in post-disaster satellite imagery using transfer learning with a fine-tuning technique to improve model adaptation across different disaster characteristics. Damage levels are mapped into four classes, namely No Damage, Minor Damage, Major Damage, and Destroyed. Swin Transformer is employed as the encoder due to its ability to efficiently extract local and global hierarchical features through the shifted window attention mechanism, while UPerNet serves as the decoder to aggregate multi-scale features and produce a more detailed per-pixel segmentation map. To support this implementation, the study utilizes post-disaster satellite images from six disaster events, with stages encompassing data collection, image preprocessing, base model training on the source domain, and fine-tuning on the target domain. Training on the source domain yielded a base model with an mIoU of 0.4537 and an F1-Score of 0.6185. This base model was then used as the initial weights for fine-tuning on the target domain through several data quantity scenarios, namely zero shot, 25% data, 50% data, and 100% data, to assess the effect of target domain data quantity on model performance. Model performance was evaluated using mIoU, F1-Score, and accuracy metrics. The results demonstrate that the transfer learning method effectively improves model adaptation across different disaster characteristics. In the zero shot scenario, the model achieved an mIoU of 0.3861 and an F1-Score of 0.5270. After fine-tuning, performance improved consistently, reaching an mIoU of 0.4831 and an F1-Score of 0.6349 in the 100% target domain data scenario.

Item Type:	Thesis (Other)
Uncontrolled Keywords:	Segmentasi Semantik, Kerusakan Bangunan, Swin Transformer, Transfer Learning, Citra Satelit, Semantic Segmentation, Building Damage, Swin Transformer, Transfer Learning, Satellite Imagery
Subjects:	Q Science Q Science > QA Mathematics Q Science > QA Mathematics > QA336 Artificial Intelligence Q Science > QA Mathematics > QA76.87 Neural networks (Computer Science)
Divisions:	Faculty of Science and Data Analytics (SCIENTICS) > Mathematics > 44201-(S1) Undergraduate Thesis
Depositing User:	Syifa Hayyina Farid
Date Deposited:	27 Jul 2026 02:16
Last Modified:	27 Jul 2026 02:16
URI:	http://repository.its.ac.id/id/eprint/137435

Actions (login required)

View Item