Enhancing geotechnical damage detection with deep learning: a convolutional neural network approach.

Area of Science:

• Geotechnical Engineering
• Computer Vision
• Artificial Intelligence

Background:

• Geodynamic events like landslides pose significant risks, causing environmental and human losses.
• Traditional visual inspections of geotechnical structures are often unsafe and impractical due to site conditions.
• Computational methods offer a feasible alternative for rapid and secure evaluations of structural integrity.

Purpose of the Study:

• To develop and validate a computer vision model for detecting surface defects in geotechnical structures.
• To reduce reliance on manual, on-site inspections by leveraging unmanned aerial vehicles (UAVs) and mobile devices.
• To address the need for specialized algorithms in geotechnical engineering, considering limited datasets and image redundancy.

Main Methods:

• Acquisition of surface failure indicator images from slopes using UAVs and mobile devices.
• Development of a custom, low-complexity convolutional neural network (CNN) for binary image classification.
• Training and testing the CNN model to distinguish between 'damage' and 'intact' states of geotechnical surfaces.

Main Results:

• The proposed CNN model achieved a high average accuracy of 94.26% in detecting surface defects.
• The model demonstrated excellent performance with an AUC score of 0.99 from the ROC curve.
• Confusion matrix analysis confirmed the model's strong capability in classifying damaged versus intact geotechnical surfaces.

Conclusions:

• The developed AI model effectively identifies failure indicators on geotechnical surfaces, enabling early detection of potential disasters.
• This approach enhances safety and efficiency in monitoring critical infrastructure, preventing catastrophic failures.
• Early detection facilitates timely maintenance and alerts, crucial for soil integrity and surrounding structures.