Predicting earthquake building damage grades using MLP with class imbalance handling global sensitivity analysis and multi method explainability

Accurate estimation of the damage grades of buildings after an earthquake is crucial for proper disaster management, infrastructure planning, and policy formulation. Although machine learning techniques are increasingly used in seismic damage grading, three issues have been inadequately explored in the process: dealing with class imbalances in the post-disaster data, quantifying the importance of features, and model interpretation. In this research, a framework for the estimation of the damage grades of buildings following earthquakes using Multilayer Perceptron (MLP) classifier is proposed on the basis of a huge dataset of buildings damaged in the post-2015 Gorkha earthquake in Nepal, consisting of 762,106 instances distributed into five classes. A total of ten resampling algorithms are compared to overcome the problem of imbalance, with the ADASYN technique performing the best in terms of accuracy (88.68%), whereas the SMOTE-LOF and SMOTified-GAN have performed better in generating the maximum macro F1-score (0.87). The global sensitivity analysis through Morris screening and Sobol variance analysis reveals that post- and pre-earthquake building heights and building ages are the two significant factors affecting predictions, as indicated by the first-order Sobol index values close to zero, indicating that predictions depend heavily on interactions between different features rather than individual features. Three explainability methods, namely SHAP, LIME, and ELI5, are employed to provide global and local explanations for the model, showing consistently that the technical solution proposed, condition post-earthquake, and post-earthquake floors are significant predictors in all three approaches.