A structural damage ranking using ConvNeXt for post-earthquake image classification

Abstract A structural damage classifier utilizing ConvNeXt, a state-of-the-art deep convolutional and residual neural network, is proposed for the automatic evaluation of post-earthquake images. The classifier distinguishes between two types of damage: structural and nonstructural. To achieve high accuracy and reliability, transfer learning was used to fine-tune the ConvNeXt model with our dataset. The model was trained on 9,645 labeled images from 1,789 reinforced concrete buildings affected by the Elazığ earthquake on January 24, 2020. To enhance detection performance while minimizing training costs, various transfer learning strategies, data augmentation, and regularization techniques were implemented. The final model was tested on images of reinforced concrete buildings taken after the Kahramanmaraş Earthquake on February 6, 2023. The results demonstrated consistent classification performance aligned with domain knowledge, along with strong generalization on the Kahramanmaraş dataset, highlighting the effectiveness of ConvNeXt in post-earthquake damage assessment.