Researchers from Fondazione Eucentre and the Artificial Vision Laboratory of the Department of Electrical, Computer and Biomedical Engineering at the University of Pavia recently published an innovative study in the Engineering Applications of Artificial Intelligence journal. The paper, entitled Post-earthquake structural damage detection with tunable semi-synthetic image generation, is the result of the joint work of Piercarlo Dondi, Alessio Gullotti, Michele Inchingolo, Ilaria Senaldi, Chiara Casarotti, Luca Lombardi and Marco Piastra.
The work is part of a context of increasing use of Artificial Intelligence for emergency management, with particular attention to the monitoring of structures affected by seismic events. The research proposes an advanced methodology for automatic structural damage detection through the combined use of Deep Learning techniques and the generation of semi-synthetic images, overcoming one of the main limitations of artificial intelligence models: the scarcity of suitable datasets.
The importance of early detection of post-earthquake damage
One of the most critical tasks after an earthquake is assessing the structural safety of buildings and infrastructure to determine whether they can be used without risk. Currently, engineering experts carry out this activity through visual inspections, a process that, although highly skilled, is extremely time-consuming and resource-intensive. Integrating Artificial Intelligence into this process could represent a turning point, allowing large amounts of data to be analysed quickly, reducing the engineers’ workload and improving the analysis’s timeliness.
Drones (Unmanned Aircraft Systems – UAS) are already widespread in post-seismic inspection operations due to their ability to capture images and videos from viewpoints inaccessible to humans, such as during bridge inspections, an activity that the Eucentre Foundation has been carrying out for some time. However, the main challenge remains the automatic and reliable analysis and interpretation of visual data.
An innovative method for creating realistic datasets
One of the main obstacles in developing automatic damage detection systems is the lack of sufficiently large and well-annotated datasets of authentic images. The number of pictures of post-seismic structural damage is limited, and their variability can make it challenging to train artificial intelligence algorithms capable of operating accurately on new scenarios.
To address this issue, the study’s authors developed an innovative method of semi-synthetic image generation. By creating three-dimensional (3D) models of actual buildings and bridges, virtual damage can be controlled by simulating different types of structural damage, such as cracks, spalling, corrosion, and leaching. These semi-synthetic images are then used to enhance the training of convolutional neural networks (DCNN), increasing the system’s ability to accurately recognise damage in authentic photos.
Validation and study results
To test the method’s effectiveness, the researchers used the IDEA dataset (Image Database for Earthquake Damage Annotation), developed by the Eucentre Foundation. The dataset includes images acquired during post-seismic reconnaissance conducted following the L’Aquila (2009), Emilia (2012), and Central Italy (2016-2017) earthquakes.
The study’s results show that integrating semi-synthetic images into the training process significantly improves the damage detection system’s performance. In particular, the developed model achieved higher accuracy than models trained solely on real images, demonstrating the effectiveness of the proposed strategy.
In addition, the system was integrated into an analysis platform for automatically processing videos captured by drones. The platform allows experts to trace the evolution of damage in video sequences and automatically generate summary reports, facilitating their work and speeding up emergency operations.
Future perspectives
The study represents a significant step forward in post-seismic structural monitoring and paves the way for further developments. Future research developments, which are still ongoing at the Foundation, include:
– The extension of the method to additional types of structural damage, such as partial collapse of buildings and the presence of rubble.
– Integration with advanced Computer Vision algorithms to improve the system’s ability to distinguish between different levels of damage.
– The algorithm is optimised to allow real-time processing directly on the drones, reducing the need for post-processing on external servers.
🔗 To read more, the full article is available
The dataset will be published soon. We will inform you through our digital channels regarding IDEA.