[Paper Review] Post-Earthquake Assessment of Buildings Using Deep Learning

Dhananjay Nahata, Harish Kumar Mulchandani+2 more|arXiv (Cornell University)|Jul 18, 2019

Infrastructure Maintenance and Monitoring17 references4 citations

TL;DR

This paper proposes a CNN-based deep learning model for automated post-earthquake building damage assessment using aerial or ground-level images. Trained on 1,200 images categorized into four damage levels—no damage, minor, major, and collapse—the VGG16 transfer learning model achieved 97.85% training accuracy and 89.38% validation accuracy with a learning rate of 1e-5, demonstrating strong real-time potential for disaster response.

ABSTRACT

Classification of the extent of damage suffered by a building in a seismic event is crucial from the safety perspective and repairing work. In this study, authors have proposed a CNN based autonomous damage detection model. Over 1200 images of different types of buildings-1000 for training and 200 for testing classified into 4 categories according to the extent of damage suffered. Categories are namely, no damage, minor damage, major damage, and collapse. Trained network tested by the application of various algorithms with different learning rates. The most optimum results were obtained on the application of VGG16 transfer learning model with a learning rate of 1e-5 as it gave a training accuracy of 97.85% and validation accuracy of up to 89.38%. The model developed has real-time application in the event of an earthquake.

Motivation & Objective

• To develop an automated, deep learning-based system for classifying building damage after earthquakes to enhance safety and repair efficiency.
• To address the limitations of manual damage assessment, which is time-consuming and resource-intensive in post-disaster scenarios.
• To evaluate the performance of transfer learning models, particularly VGG16, on a diverse dataset of building images with varying damage levels.
• To identify optimal hyperparameters, such as learning rate, for maximizing classification accuracy in real-world deployment.
• To enable real-time, scalable damage assessment using deep learning for emergency response and urban resilience planning.

Proposed method

• Collected and curated a dataset of 1,200 building images—1,000 for training and 200 for testing—categorized into four damage classes: no damage, minor, major, and collapse.
• Employed transfer learning with the pre-trained VGG16 architecture to leverage learned features from ImageNet for improved generalization on the damage classification task.
• Fine-tuned the VGG16 model using different learning rates (e.g., 1e-5) to optimize performance on the custom dataset.
• Applied data augmentation and normalization techniques to improve model robustness and generalization across diverse image conditions.
• Trained the model using stochastic gradient descent with cross-entropy loss to minimize classification error.
• Evaluated model performance using standard metrics: training and validation accuracy, with results reported across all four damage categories.

Experimental results

Research questions

• RQ1 Can a deep learning model effectively classify building damage into four distinct categories using post-earthquake images?
• RQ2 How does transfer learning with VGG16 compare to training from scratch in terms of accuracy and convergence speed on a limited damage dataset?
• RQ3 What learning rate optimizes the performance of a fine-tuned VGG16 model for building damage classification?
• RQ4 To what extent can the model generalize to real-world post-disaster scenarios with variable image quality and lighting?
• RQ5 Can the model achieve high validation accuracy while remaining computationally efficient for real-time deployment in emergency response?

Key findings

• The VGG16 transfer learning model achieved a training accuracy of 97.85% and a validation accuracy of 89.38% when fine-tuned with a learning rate of 1e-5.
• The model demonstrated strong generalization on unseen test images, indicating robustness to variations in building types and image conditions.
• Among the tested learning rates, 1e-5 yielded the highest validation accuracy, suggesting optimal convergence for this dataset and architecture.
• The model successfully distinguished between all four damage categories, including the critical distinction between major damage and collapse.
• The use of transfer learning significantly improved performance compared to training from scratch, as evidenced by high accuracy with relatively fewer training epochs.
• The results confirm the feasibility of deploying deep learning models for real-time, automated building damage assessment in post-earthquake scenarios.