Building damage level classification using deep learning: a CNN-based approach for post-earthquake structural assessment

In the aftermath of major earthquakes, a rapid and accurate structural damage assessment is crucial for emergency response and recovery. Convolutional Neural Networks (CNNs) have emerged as effective tools for automating this process, offering standardized evaluations that complement traditional visual inspections. This study explores the use of the VGG16 archi- tecture for post-earthquake damage classification, leveraging transfer learning and data aug- mentation techniques to enhance accuracy. The dataset comprises 5,000 RGB images sourced from the PHI-Net dataset and the INGV DFM database, categorized into four damage lev- els. Through extensive pre-processing and augmentation, VGG16 achieved a test accuracy of 89.33%, with high precision and recall for undamaged and severe damage classes. How- ever, distinguishing minor damage remains still challenging. These findings highlight CNNs’ potential in automating structural damage assessment, supporting more efficient post-disaster decision-making.