Deep Learning Enhances Building Risk Assessment

By Nidhi DhullReviewed by Susha Cheriyedath, M.Sc.Sep 16 2024

A recent article published in Discover Applied Sciences proposed using computer vision deep learning models to automatically classify building characteristics and create large-scale (city or region) exposure models for risk assessment. The data extracted from the best convolutional neural network (CNN)-based model was compared with traditional ground data.

Background

Accurate and up-to-date building information is critical for reducing the impact of natural hazards. Effective hazard impact assessments require detailed exposure models that account for a building’s location, value, occupancy, and construction attributes.

Traditionally, exposure models have been built using data from cadastral records, national housing censuses, rapid visual screenings, or approximations based on demographic or socio-economic data. However, these approaches often lack specific information about individual buildings and their exact locations, leading to greater uncertainty when classifying a building’s vulnerability to natural hazards.

Recent advancements in building imagery have opened up new possibilities, allowing the development of more precise datasets at the urban level for risk assessments. By combining street-level imagery with machine learning algorithms, it is now possible to automatically and accurately identify specific building characteristics. This study leverages convolutional neural networks (CNNs) to create exposure models that predict attributes such as a building’s construction material and age.

Methods

The dataset used in this study was extracted from the Alvalade region (Lisbon, Portugal). The building footprints retrieved from the region’s OpenStreetMap were further verified using the quantum geographic information system (QGIS) platform. Moreover, street-level images (2670 building images) of the façade of each building were collected for each building footprint using the Google Street View (GSV) software.

Additional data (4085 pictures) for training, validating, and assessing the deep learning models was collected through fieldwork. Up to 3 photos from different angles of each building’s façade and other information such as construction material, construction year, buildings’ rehabilitation, etc., were captured.

An earthquake engineer helped label the collected data for quality control. Subsequently, images without a façade label were discarded. The remaining 4239 images were classified as A (construction material and the range of the number of floors), B (a finer distribution of the number of floors and construction material), and C (construction epoch).

The dataset was split into training (80 %, 3391 images) and test (20 %, 848 images) subsets. The training subset was further divided into a validation subset (20 %, 678 images). This data was used to train and test different CNN-based models (ResNet50V2, InceptionResNetV2, NASNetLarge, Xception, InceptionV3, and DenseNet201).

The performance of these models in classifying different buildings’ characteristics was compared. Subsequently, the risk assessment results based on data extracted from the best CNN-based model were compared against traditional ground data results.

Results and Discussion

The models trained using transfer learning and fine-tuning showed that Xception performed best for configurations A and B, while DenseNet201 outperformed for configuration C. Interestingly, the highest accuracy was observed in configuration A and the lowest in configuration C. This indicates that construction material and building height (number of floors) were more reliable classification parameters than construction epoch.

The Xception model achieved high precision and recall for masonry buildings. As shown in the confusion matrix, only 11.9 % of masonry buildings were misclassified as concrete, which would result in a slight underestimation of damage during an earthquake. Nevertheless, recall values indicated that over 89 % of masonry buildings were correctly identified, regardless of their height, ensuring that the exposure models did not significantly underestimate risk on a broad scale.

For concrete buildings, the model's precision and recall were even higher than for masonry. Although some masonry buildings were misclassified as concrete, leading to risk underestimation, the misclassifications primarily occurred between classes with similar vulnerabilities. Consequently, this margin of error was considered acceptable and did not compromise the overall risk assessment.

Two exposure models were developed to predict seismic risk in the Alvalade region, one based on field data collected by civil engineers and the other using building characteristics automatically classified by CNN models. When assessing seismic risk from a hypothetical onshore event with a moment magnitude of 6.0, both models predicted nearly identical impacts despite the 12 % building misclassification rate.

Conclusion and Future Prospects

The researchers successfully demonstrated the effectiveness of deep CNNs in automating the classification of building stock for exposure modeling using Google Street View (GSV) images. Of the three tested sets of building characteristics, the Xception model achieved the highest classification accuracy, exceeding 80 %.

The exposure model using automatically classified data performed on par with traditional ground data-based models in terms of accuracy while significantly reducing time and costs. Incorporating additional data sources could further improve the accuracy of the CNN-based model.

Looking ahead, the researchers plan to develop earthquake exposure models using these algorithms to automatically classify building stock. They also intend to apply the trained models to other regions of Lisbon to evaluate their generalization capabilities.

Journal Reference

Gouveia, F., Silva, V., Lopes, J., Moreira, R. S., Torres, J. M., & Guerreiro, M. S. (2024). Automated identification of building features with deep learning for risk analysis. Discover Applied Sciences, 6(9). DOI: 10.1007/s42452-024-06070-2, https://link.springer.com/article/10.1007/s42452-024-06070-2

Disclaimer: The views expressed here are those of the author expressed in their private capacity and do not necessarily represent the views of AZoM.com Limited T/A AZoNetwork the owner and operator of this website. This disclaimer forms part of the Terms and conditions of use of this website.