Exploratory Image-Based Defect Analysis of Sikatak Bridge Using YOLO Segmentation Models with Complementary Terrestrial Laser Scanning

Abstract

This study presents an exploratory image-based defect analysis and complementary terrestrial laser scanning (TLS) documentation of Sikatak Bridge, a steel arch bridge with concrete components located in Semarang, Central Java, Indonesia. A Leica RTC360 produced a colourised point cloud, while smartphone photographs illustrated visible surface defects. Training records for YOLOv9c-seg and a run labelled YOLOv11seg were examined using separate bounding-box and mask validation metrics. YOLOv9c-seg stopped after 81 of 150 configured epochs because the monitored validation fitness did not improve for 30 epochs; its selected checkpoint occurred at epoch 51. The V11 CSV contained 150 epochs, but a matching complete notebook was unavailable. Peak-mAP50-95 validation rows gave box values of 0.384 and 0.425 and mask values of 0.371 and 0.426 for the V9 and V11 files, respectively. These values are descriptive log summaries, not evidence of model superiority. The YOLOv9 loader rejected 924 of 1,013 validation images because of annotation incompatibilities, leaving 89 evaluated images; the original class mapping and the V11 evaluation membership remain unresolved. The reported 42 field photographs could not be traced to the downloaded training corpus. Consequently, the study does not establish reliable four-class accuracy, independent-test generalisation, or comparative computational efficiency. The point-cloud views document bridge geometry qualitatively, without validated defect localisation or dimensional accuracy. Dataset reconciliation, annotation verification and matched independent testing are required before operational use.