Identifikasi Tingkat Risiko Gangguan Kabel Udara di Area Vegetasi menggunakan Segmentasi Multi-class EfficientViT U-Net dan Klasifikasi Rule-based

Jannah, Alissa Velia Royhatul (2026) Identifikasi Tingkat Risiko Gangguan Kabel Udara di Area Vegetasi menggunakan Segmentasi Multi-class EfficientViT U-Net dan Klasifikasi Rule-based. Masters thesis, Institut Teknologi Sepuluh Nopember.

Text 6025241027-Master_Thesis.pdf - Accepted Version Restricted to Repository staff only Download (5MB) | Request a copy

Abstract

Infrastruktur kabel udara terdiri dari kabel listrik dan kabel telekomunikasi, yang biasanya berdampingan dengan area vegetasi. Vegetasi yang tumbuh di luar batas aman berpotensi menimbulkan bahaya keselamatan dan keandalan kabel udara, mulai dari terputusnya komunikasi digital hingga pemadaman listrik total. Selain berdampak pada kehidupan sehari-hari, juga berdampak pada kerugian finansial yang dialami oleh perusahaan penyedia infrastruktur. Oleh karena itu, identifikasi terhadap potensi gangguan kabel udara sangat penting untuk dilakukan.

Penelitian ini menganalisis pendekatan yang menggabungkan teknik segmentasi dan klasifikasi untuk identifikasi risiko gangguan kabel udara. Pertama, arsitektur U-Net digunakan sebagai baseline segmentasi citra dengan modifikasi pada bagian encoder, menggunakan EfficientNet untuk meningkatkan representasi feature extraction dengan teknik compound scaling dalam meningkatkan resolusi, kedalaman, dan lebar jaringan dasar CNN. Kemudian, modifikasi pada bagian bottleneck, menggunakan Vision Transformer (ViT) dengan komponen inti Multi-head Self-Attention (MSA) untuk menangkap konteks global untuk mengatasi kelemahan CNN konvensional yang hanya unggul dalam fitur lokal. Kedua, dilakukan klasifikasi berbasis aturan berdasarkan pengetahuan dari ahli kabel udara dengan menghitung mask area dari hasil segmentasi. Klasifikasi berbasis aturan menghasilkan tiga jenis risiko, “High” berdasarkan area tumpang tindih antara kabel udara dan vegetasi, “Low” berdasarkan area kabel udara dan vegetasi yang tidak tumpang tindih, “Undefined” berdasarkan area yang terdeteksi bukan objek kabel udara dan vegetasi.

Eksperimen dilakukan menggunakan dataset publik VEPL yang dianotasi ulang menjadi NEW-VEPL-4 untuk kebutuhan segmentasi multi-class. Arsitektur segmentasi multi-class EfficientNetB5 + ViT yang diusulkan berhasil mengungguli model baseline dengan nilai rata-rata tertinggi pada seluruh parameter evaluasi, yakni IoU sebesar 0,894, Dice Coefficient sebesar 0,941, presisi sebesar 0,943, dan recall sebesar 0,939. Sementara itu, pengujian terhadap evaluasi rule-based mencapai nilai rata-rata akurasi sebesar 0,940, presisi sebesar 0,934, recall sebesar 0,752, dan F1-score sebesar 0,800.
=====================================================================================================================================
Overhead cable infrastructure consists of power and telecommunications cables, which are typically located adjacent to vegetated areas. Vegetation growing beyond safe limits has the potential to pose safety and reliability hazards to overhead cables, ranging from disruptions in digital communications to total power outages.  In addition to impacting daily life, this also results in financial losses for infrastructure providers. Therefore, identifying potential overhead cable disruptions is crucial.

This study analyzes an approach that combines segmentation and classification techniques to identify the risk of overhead cable faults. First, the U-Net architecture is used as a baseline for image segmentation with modifications to the encoder section, utilizing EfficientNet to enhance feature extraction by applying compound scaling to increase the resolution, depth, and width of the base CNN network. Next, modifications were made to the bottleneck layer by incorporating a Vision Transformer (ViT) with a Multi-head Self-Attention (MSA) core to capture global context, thereby addressing the limitation of conventional CNNs, which excel only in local features. Second, rule-based classification was performed using expertise from overhead power line specialists by calculating the masked area from the segmentation results. Rule-based classification yields three risk categories: “High,” based on the overlap area between overhead cables and vegetation; “Low,” based on areas where overhead cables and vegetation do not overlap; and “Undefined,” based on areas where neither overhead cables nor vegetation are detected.

The experiments were conducted using the public VEPL dataset, which was re-annotated as NEW-VEPL-4 for multi-class segmentation. The proposed EfficientNetB5 + ViT multi-class segmentation architecture outperformed the baseline model with the highest average scores across all evaluation metrics: an IoU of 0.894, a Dice Coefficient of 0.941, a precision of 0.943, and a recall of 0.939. Meanwhile, testing using rule-based evaluation yielded average accuracy of 0.940, precision of 0.934, recall of 0.752, and an F1-score of 0.800.

Item Type:	Thesis (Masters)
Uncontrolled Keywords:	Gangguan Kabel Udara, Klasifikasi Berbasis Aturan, Segmentasi Citra Vegetasi, Segmentasi Multikelas, Aerial Cable Interference, Rule-Based Classification, Vegetation Image Segmentation, Multi-class segmentation.
Subjects:	Q Science > QA Mathematics > QA336 Artificial Intelligence Q Science > QA Mathematics > QA76.76.E95 Expert systems Q Science > QA Mathematics > QA76.87 Neural networks (Computer Science)
Divisions:	Faculty of Industrial Technology > Informatics Engineering > 55101-(S2) Master Thesis
Depositing User:	Alissa Velia Royhatul Jannah
Date Deposited:	31 Jul 2026 01:56
Last Modified:	31 Jul 2026 01:56
URI:	http://repository.its.ac.id/id/eprint/139766

Actions (login required)

View Item