POST-EARTHQUAKE DAMAGE STATE CLASSIFICATION USING VGG-16 CONVOLUTIONAL NEURAL NETWORK

Abstract

In the aftermath of an earthquake, rapid and accurate assessment of structural damage is crucial for effective response and recovery efforts. Traditional methods often involve time-consuming and manual inspections, limiting the speed and scale of damage assessment. The use of emerging technologies such as Machine Learning can significantly reduce the resources needed to accurately and rapidly assess the damage dealt to structures. By harnessing the power of deep learning, an automated approach is proposed to streamline this process and provide timely and objective damage detection. This study investigates the use of VGG-16 convolutional neural network, a widely adopted deep learning architecture known for its effectiveness in image classification tasks, for post-earthquake damage assessment. The model is trained using images from a database of previous earthquakes in the Philippines and the extensive PEER Hub Image-Net (-Net) dataset. These datasets encompass a wide array of annotated images depicting earthquake-damaged structures. The structures in the photos were classified by qualified engineering practitioners according to the damage states specified in FEMA HAZUS MH2.1: no damage, slight damage, moderate damage, extensive damage, and complete damage. Results show that the proposed VGG-16 convolutional neural network achieved a promising accuracy of 65.1% in accurately classifying the damage states of earthquake-damaged structures. The model's performance indicates its ability to effectively differentiate between various damage states, ranging from no damage to complete damage. Future work involves the further tuning of the hyperparameters and training the model with additional post-earthquake photos. This work affirms the potential of deep learning methods in augmenting the efforts in post-earthquake reconnaissance surveys.