Deep learning for bridge load capacity estimation in post-disaster and -conflict zones

Area of Science:

• Civil Engineering
• Computer Science
• Artificial Intelligence

Background:

• Post-disaster and post-conflict regions often lack essential transportation infrastructure data, impeding mobility and reconstruction.
• Deteriorating bridges necessitate accurate quantification of load characteristics for maintenance and asset management.
• Manual bridge assessment is time-consuming, costly, and requires specialized expertise, especially in challenging environments.

Purpose of the Study:

• To propose and evaluate a deep learning approach for estimating bridge load carrying capacity using crowdsourced images.
• To address the data scarcity issue for transportation infrastructure in critical regions.
• To provide a scalable and efficient method for assessing bridge structural integrity.

Main Methods:

• A convolutional neural network (CNN) architecture was developed and trained on a dataset of over 6000 bridges.
• The study analyzed the impact of dataset variations (e.g., image quality, class intervals) on prediction performance.
• Multiclass classification was converted to binary classification for practical field application.

Main Results:

• The deep learning model demonstrated the ability to estimate bridge load carrying capacity from crowdsourced images.
• Significant variations within the dataset were identified and their effects on prediction metrics (accuracy, precision, recall, F1 score) were quantified.
• Binary classification optimization yielded promising performance for field use.

Conclusions:

• Deep learning offers a viable solution for estimating bridge load capacity, particularly in data-scarce regions.
• Crowdsourced imagery can be effectively leveraged for infrastructure assessment using AI.
• The developed method has the potential to significantly improve bridge maintenance strategies and asset database accuracy.