Construction of Asbestos Slate Deep-Learning Training-Data Model Based on Drone Images

Area of Science:

• Environmental Science
• Computer Science
• Materials Science

Background:

• Visual inspection of asbestos roof slate poses safety risks and incurs high costs.
• Drone technology and deep learning offer potential solutions for efficient and safe detection.
• Accurate identification of asbestos-containing materials is crucial for public health and safety.

Purpose of the Study:

• To develop a comprehensive deep-learning model for accurate asbestos roof slate detection using drone imagery.
• To evaluate the effectiveness of combining supervised and unsupervised classification techniques.
• To improve the speed and reduce the cost of asbestos detection.

Main Methods:

• A deep-learning model was developed using supervised and unsupervised classification.
• High-resolution drone imagery was captured at a low altitude (100 m) for a ground sampling distance of 3 cm/pixel.
• The model was trained and validated using diverse image data under various conditions (lighting, weather, angles).

Main Results:

• The initial model achieved high accuracy, with only 12 misclassifications out of 475 images.
• Post-classification adjustments and model retraining led to precise classification of all images.
• The combined classification approach demonstrated significant improvements in detection accuracy.

Conclusions:

• Supervised and unsupervised classification methods can be effectively combined to enhance deep-learning model accuracy for asbestos roof slate detection.
• Drone-based detection with advanced AI offers a safer, faster, and more cost-effective alternative to traditional methods.
• The developed model shows high potential for real-world application in asbestos management and safety.