Classification Of Superstructure Damage In School Buildings In Nusa Penida Bali Using YOLO V7

Abstract

Structural damage in school buildings poses significant risks to safety and education quality, particularly in remote areas with limited maintenance resources. This study develops a YOLOv7-based model to detect building pillars and classify structural damages, focusing on school buildings in Nusa Penida, Bali. A dataset of 156 images, derived from an initial 521 images collected during field visits, was curated to include both damaged and intact pillars. Preprocessing and augmentation techniques, including resizing and rotation, were applied to optimize the dataset. Training was conducted over 55 epochs using Google Colab with a T4 GPU, incorporating parameter tuning to address dataset imbalance. Confidence thresholds were set at 0.7 for pillars and 0.2 for rebar detection to enhance sensitivity to underrepresented damage classes. Evaluation metrics, including the F1-score and confusion matrix, confirmed the model’s accuracy and robustness in detecting and classifying structural damages. The results demonstrate the model's potential for real-world applications in damage assessment, particularly in resource-limited settings. Future research should focus on expanding datasets, incorporating multi-class classification, and integrating real-time detection and drone-based imagery to enhance scalability and efficiency. This work contributes to developing efficient, AI-driven solutions for structural health monitoring in critical infrastructure.