Transforming Building Risk Assessment with Deep Learning

In recent years, deep learning has significantly transformed the landscape of building risk assessment. By leveraging advanced computer vision models, researchers are now able to automatically classify building characteristics, thereby creating comprehensive exposure models that enhance risk evaluation. This innovative approach was highlighted in a study that utilized convolutional neural network (CNN)-based models, comparing their data with traditional ground data to assess their effectiveness.

Accurate information about buildings is crucial for minimizing the impact of natural disasters. Effective hazard assessments depend on detailed exposure models that account for a building's location, value, occupancy, and construction details. Traditional methods, which often rely on cadastral records, national housing censuses, and rapid visual screenings, may not provide specific information about individual buildings, leading to uncertainties in evaluating their vulnerability to natural hazards.

Advancements in building imagery have paved the way for more precise urban-level datasets used in risk assessments. By integrating street-level imagery with machine learning algorithms, it is now feasible to automatically identify specific building characteristics. The study in question employed CNNs to develop exposure models that predict attributes such as construction material and age.

Methodology

The research was conducted in the Alvalade region of Lisbon, Portugal. Building footprints from OpenStreetMap were verified using the quantum geographic information system (QGIS), and street-level images of building façades were collected through Google Street View. Additional data for training and validating the deep learning models were obtained via fieldwork, which included images and information on construction materials, year, and rehabilitation.

An earthquake engineer was responsible for labeling the data to ensure quality control, discarding images that lacked façade labels. The dataset was organized into three configurations based on construction material, number of floors, and construction epoch. The data was divided into training and test subsets, and various CNN-based models were trained and tested, including ResNet50V2, InceptionResNetV2, NASNetLarge, Xception, InceptionV3, and DenseNet201.

Results and Discussion

The study revealed that the Xception model performed optimally for configurations A and B, while DenseNet201 excelled in configuration C. The highest accuracy was noted in configuration A, indicating that construction material and building height are more reliable classification parameters than construction epoch.

The Xception model showed high precision and recall for masonry buildings, with a minimal misclassification rate. For concrete buildings, the model's precision and recall were even higher. Despite some misclassifications, the overall risk assessment remained accurate, with acceptable margins of error.

Two exposure models were developed to predict seismic risk in the Alvalade region: one based on field data and the other using CNN-classified building characteristics. Both models predicted nearly identical impacts from a hypothetical seismic event, despite a 12% building misclassification rate.

Conclusion and Future Prospects

The study successfully demonstrated the potential of deep CNNs in automating the classification of building stock for exposure modeling using Google Street View images. The Xception model achieved over 80% classification accuracy, comparable to traditional ground data-based models, while reducing time and costs. Future improvements could involve incorporating additional data sources to enhance the CNN-based model's accuracy.

Researchers plan to develop earthquake exposure models using these algorithms and apply the trained models to other regions of Lisbon to evaluate their generalization capabilities.