Wind turbine blade damage detection — YOLOv8s
Fine-tuned from yolov8s.pt (COCO) on the WTBD dataset.
Classes
craze, corrosion, surface_injure, thunderstrike, crack, hide_craze
Class indices are stored in the checkpoint as model.names — read them from
there rather than hardcoding, so a reordered data.yaml cannot mislabel
predictions.
Training
- 100 epochs, imgsz=1024, seed=42, batch auto
- Ultralytics 8.3.40
Metrics (held-out test split)
mAP50: 0.805 — mAP50-95: 0.503
Data source
Ji, L., Cheng, J., Wu, S. Multiclass Dataset for Intelligent Detection of Wind Turbine Blade Defects Using Drone Imagery. Scientific Data 13, 396 (2026). https://doi.org/10.1038/s41597-026-06762-x