Search

• A novel multi-class factorization machine approach with deep neural network is developed. • It performs post-disaster building damage assessment in an automatic and data-driven manner. • A total of 39,352 buildings in 2015 Napa earthquake are taken as a case study for demonstration. • It improves the classification performance over many other popular machine learning methods. • The well-trained model can significantly reduce the burden in earthquake field investigations. This paper develops a hybrid neural network architecture named multi-class factorization machine with deep neural network (multi-FMDNN) to fuse multi-source information for the automatic post-earthquake building damage evaluation. The novel algorithm is a combination of the factorization machine (FM) and the deep neural network (DNN), which adopts the one-vs-all strategy to fuse results from multiple base classifiers. 39,352 buildings affected by the 2015 Nepal earthquake are taken as a case study to validate the effectiveness of the proposed multi-FMDNN. Experimental results confirm that the proposed model outperforms over many other popular machine learning methods due to the powerful feature learning ability, ultimately reaching an overall accuracy, macro F1-score, and weighted F1-score in the value of 0.703, 0.737, and 0.702, respectively. Features associated with building structural characteristics are found to contribute more to classifying damage grades precisely. Besides, data preprocessing for data cleaning, encoding, and transformation is a necessary step to bring additional performance enhancement. For significance in the knowledge aspect, a novel multi-FMDNN algorithm is developed, which is superior in extracting both low- and high-order feature representation automatically from large volumes of destroyed buildings-related data and learning the optimal feature interactions simultaneously to pursue more accurate classification. For significance in the application aspect, the predicted results provide deep insights into a better understanding of the building vulnerability in seismic areas and inform data-driven decisions in disaster relief efforts. A promising future scope is to make full use of the available pre-event data along with some post-event data, which is possible to return fairly promising predictions and reduce the burden in earthquake field investigations for rapid responses. In future work, advanced techniques associated with data augment, hyperparameter optimization, and others will be implemented to constantly improve the overall accuracy and generalizability of the prediction model. [ABSTRACT FROM AUTHOR]