Pengembangan Model Re-Identifikasi Penyu Menggunakan Vision Transformer

Mahmudi, Sulthan Daffa Arif (2026) Pengembangan Model Re-Identifikasi Penyu Menggunakan Vision Transformer. Other thesis, Institut Teknologi Sepuluh Nopember.

Text 5024211004-Undergraduate_Thesis.pdf - Accepted Version Restricted to Repository staff only Download (13MB) | Request a copy

Abstract

Penyu laut merupakan satwa dilindungi yang populasinya terus menurun akibat aktivitas manusia seperti perburuan ilegal dan kerusakan habitat. Upaya konservasi memerlukan sistem identifikasi individu yang akurat, sementara metode konvensional seperti satellite tagging berbiaya tinggi, berisiko rusak, dan bersifat invasif. Penelitian ini mengembangkan model re-identifikasi (Re-ID) penyu berbasis citra kepala menggunakan arsitektur Vision Transformer (ViT) pra-latih DINOv3 sebagai alternatif yang lebih efisien dan non-invasif. Model dilengkapi embedding head dan dilatih pada dataset SeaTurtleIDHeads dengan skema open-set identity split, mengombinasikan ArcFace Loss dan Online Hard Triplet Loss, Layer-wise Learning Rate Decay (LLRD), serta cosine learning rate scheduler. Sebanyak 19 parameter divariasikan secara one-factor-at-a-time untuk memetakan sensitivitasnya. Konfigurasi acuan memperoleh Rank-1 78,3% dan mAP 84,9%, sedangkan konfigurasi kombinasi mencapai Rank-1 90,4% dan mAP 93,5%. Backbone base learning rate terbukti menjadi parameter paling berpengaruh dengan kontribusi sekitar 78% dari total peningkatan, dan dampak antar-parameter bersifat sub-aditif. Pengujian generalisasi lintas-dataset pada SeaTurtleID2022 tanpa pelatihan ulang menghasilkan Rank-1 81,1% dan Rank-5 97,2%, yaitu penurunan yang terukur namun tidak katastrofik. Pada skenario identifikasi dengan penolakan identitas tidak terdaftar, Registered Rank-1 turun menjadi 55,3% (threshold global) dan 56,7% (background class) dengan false-reject rate sekitar 42%. Distribusi skor menunjukkan separabilitas yang hanya moderat (ROC-AUC 0,77–0,79), sehingga keterbatasan sistem bersumber dari mekanisme keputusan dan kalibrasi skor, bukan dari kualitas embedding.
=========================================================================================================================================
Sea turtles are protected animals whose populations continue to decline due to human activities such as illegal hunting and habitat destruction. Conservation efforts require an accurate individual identification system, while conventional methods such as satellite tagging are costly, prone to device failure, and invasive. This study develops a sea turtle re-identification (Re-ID) model based on head images using a Vision Transformer (ViT) architecture pretrained with DINOv3 as a more efficient and non-invasive alternative. The model is equipped with an embedding head and trained on the SeaTurtleIDHeads dataset under an open-set identity split, combining ArcFace Loss and Online Hard Triplet Loss, Layer-wise Learning Rate Decay (LLRD), and a cosine learning rate scheduler. Nineteen parameters were varied one-factor-at-a-time to map their sensitivity. The baseline configuration achieved Rank-1 of 78.3% and mAP of 84.9%, while the combined configuration reached Rank-1 of 90.4% and mAP of 93.5%. The backbone base learning rate proved to be the most influential parameter, contributing approximately 78% of the total improvement, and parameter effects were found to be sub-additive. Cross-dataset generalization on SeaTurtleID2022 without retraining yielded Rank-1 of 81.1% and Rank-5 of 97.2%, a measurable but non-catastrophic degradation. Under identification with rejection of unregistered identities, Registered Rank-1 dropped to 55.3% (global threshold) and 56.7% (background class), with a false-reject rate of around 42%. Score distributions showed only moderate separability (ROC-AUC 0.77–0.79), indicating that the system's limitation stems from the decision mechanism and score calibration rather than embedding quality.

Item Type:	Thesis (Other)
Uncontrolled Keywords:	Penyu Laut, Sea Turtles, Re-identification, Vision Transformer, DINOv3, Open-set
Subjects:	T Technology > T Technology (General) > T57.5 Data Processing
Divisions:	Faculty of Intelligent Electrical and Informatics Technology (ELECTICS) > Computer Engineering > 90243-(S1) Undergraduate Thesis
Depositing User:	Sulthan Daffa Arif Mahmudi
Date Deposited:	29 Jul 2026 03:32
Last Modified:	29 Jul 2026 03:32
URI:	http://repository.its.ac.id/id/eprint/138903

Actions (login required)

View Item