JAUE2024-082: Construction of Deep Learning Training Data Models for Asbestos Slate Detection Using UAV Imagery

Abstract

This study proposes a methodology for constructing a deep learning training dataset to detect asbestos slate roofing using drone imagery and image classification techniques. Aerial images of asbestos slate roofs were captured using Unmanned Aerial Vehicle (UAV) at 5 locations, collecting 475 total images. Unsupervised classification was performed on the images using ISO Cluster, dividing them into 250 classes. The classes corresponding to the ridge and groove patterns of asbestos slate were extracted and used as training data for supervised classification. Maximum likelihood classification was then applied to generate the final asbestos slate detection model. Field verification was conducted to validate the results, achieving 100% accuracy after parameter adjustments. This approach combines the strengths of unsupervised and supervised classification to efficiently generate labeled training data for deep learning asbestos detection models. The methodology enables faster, safer, and more cost-effective large-scale surveys of asbestos roofing compared to traditional visual inspection methods.