Abstract
The Shanhaiguan Great Wall, a UNESCO World Heritage Site from the Ming Dynasty, faces significant structural threats due to environmental degradation of its gray brick surfaces. Traditional manual inspection methods are labor-intensive and often dangerous. This research applies the YOLOv4 machine learning model to automate the detection of four primary damage types: chalking, plant growth, ubiquinol (efflorescence), and cracking. By training on a specialized dataset of 1,839 high-resolution images, the model achieved an overall accuracy of 85.87%. This non-destructive approach significantly improves monitoring efficiency and provides a scalable solution for the preservation of ancient masonry structures.
Introduction: The Challenge of Heritage Preservation
Preserving the Great Wall of China is a monumental task. The Shanhaiguan section is primarily composed of rammed earth and gray bricks. Over centuries, environmental stressors like wind, rain, and freeze-thaw cycles have caused significant surface damage. Identifying this damage traditionally requires experts to conduct on-site manual observations or review thousands of photos taken by Unmanned Aerial Vehicles (UAVs). This process is subjective, labor-intensive, and hazardous in difficult-to-access mountain sections. There is a critical need for an automated, objective, and non-destructive way to monitor the structural health of these ancient bricks to enable timely repairs.
Study Area and Environmental Context
The research focuses on the Plain Great Wall of Shanhaiguan. This section is particularly vulnerable due to its geographic location and architectural design.
Environmental factors play a decisive role in the degradation of the masonry. Data from the Qinhuangdao region reveals significant temperature fluctuations throughout the year, which facilitate the freeze-thaw cycles mentioned as a primary driver of brick cracking and surface chalking.
Methodology: YOLOv4 for Heritage Conservation
The researchers employed YOLOv4 (You Only Look Once, version 4), a real-time object detection framework known for balancing speed and accuracy. The architecture consists of three main components:
- Backbone (CSPDarknet53): Responsible for feature extraction from the input images.
- Neck (PANet): Fuses feature maps from different scales to ensure the model can detect both large and small damage areas.
- Head (YOLO Head): Predicts the bounding boxes and categories of the damage.
To overcome the lack of a pre-existing "Great Wall Damage" dataset, the team used transfer learning, initializing their model with weights from the COCO/VOC datasets before fine-tuning it on their specific brick imagery. They also utilized data augmentation techniques—such as random scaling, rotation, and perturbation—to improve the model's robustness against varying light conditions and shooting angles.
Dataset Development
The researchers collected a primary dataset of 1,839 high-resolution photos using mobile devices and UAVs at the Shanhaiguan site. From this pool, they curated a high-quality labeled dataset of 361 samples for training and validation, covering four categories: Chalking (107 samples), Ubiquinol (107 samples), Plant Damage (100 samples), and Brick Cracking (47 samples). The images were unified to a 512 × 512 pixel format and labeled by cultural relic experts to ensure ground-truth accuracy.
Key Insights and Results
The study yielded several critical findings regarding both the model performance and the nature of Great Wall decay:
- Architecture and Decay: The "ladder retaining wall" structure of the Great Wall, while stable, creates small horizontal planes that trap rainwater and seeds, leading to concentrated moss and plant growth.
- The Damage Pipeline: The study identifies ubiquinol (efflorescence) as a precursor to more severe damage. These soluble salts absorb moisture and eventually lead to "chalking," where the brick surface loses its structural integrity and turns to powder.
- Model Performance: The YOLOv4 model achieved an overall accuracy of 85.87%. It excelled in detecting Cracking (97.87%) and Plant Damage (97.00%) due to high contrast. Chalking achieved 91.59% accuracy, while Ubiquinol proved most challenging (64.49%) due to its subtle appearance and lack of clear boundaries.
Conclusion and Future Work
This machine learning approach validates that automated systems can replace invasive testing or manual surveys for heritage sites. While the model currently struggles with low-contrast surface defects like early-stage ubiquinol, future work will focus on expanding the training set and optimizing hyperparameters. Integrating these models directly onto UAV hardware could eventually allow for real-time, autonomous health monitoring of the entire Great Wall network.
Summary updated