Machine learning ranks building damage after Nepal’s deadly 2015 Gorkha earthquake

A machine learning framework trained on more than 762,000 buildings damaged in Nepal’s 2015 Gorkha earthquake can assign damage grades with up to 88.68 percent accuracy, and its developers say the real advance lies in how the model explains itself. The study, published in Discover Civil Engineering, pairs a Multilayer Perceptron classifier with ten resampling strategies for class imbalance, global sensitivity analysis, and three separate explainable AI methods, all applied to one of the largest post-disaster building datasets ever assembled.

The earthquake itself was catastrophic. With a magnitude of 7.8, the 2015 Gorkha event killed about 9,000 people, forced millions from their homes, and caused economic losses equal to roughly half of Nepal’s GDP. In its aftermath, the Central Bureau of Statistics, Kathmandu Living Labs, and the National Planning Commission collaborated on one of the largest post-disaster data collection efforts on record. The resulting dataset covers 762,106 building records with 39 attributes spanning structural details, materials, environmental conditions, and administrative information. Each building carries a damage grade from 1, indicating minimal damage, to 5, indicating complete devastation, assigned through field surveys.

That target variable is severely imbalanced, which is a notorious trap for classifiers. Grade 5 accounts for roughly 275,000 records while Grade 1 has only about 78,000, and a standard model can post deceptively high accuracy by simply favoring the majority class while missing the rarer, often more critical damage levels. To counter this, the researchers compared ten resampling techniques, including SMOTE, ADASYN, Tomek Links, SMOTE-Tomek, KMeans-SMOTE, SVM-SMOTE, SMOTified-GAN, ADASYN-LOF, and SMOTE-LOF, with resampling applied only to training data so the held-out test set stayed untouched.

The results showed a clear hierarchy with an instructive tension in it. ADASYN, which concentrates synthetic sample generation in boundary regions where minority instances are hardest to learn, achieved the highest accuracy at 88.68 percent. However, the best macro F1-scores, about 0.87, came from SMOTE-LOF and SMOTified-GAN, methods that filter outliers or use generative adversarial synthesis before oversampling. The baseline MLP without resampling managed 80.35 percent accuracy, and standalone Tomek Links undersampling actually reduced both accuracy and macro F1. Because no single method optimized accuracy, macro F1, and recall simultaneously, the authors argue that the choice of resampling strategy should depend on whether a response operation prioritizes overall performance or balanced detection across all five damage grades.

Beyond raw prediction, the team tackled a question most seismic damage studies skip: which features actually drive the model, and do they act alone or in combination? They applied Morris screening and Sobol variance decomposition, global sensitivity methods that distinguish direct effects from interactions without assuming anything about the model’s internal form. Morris analysis ranked post-earthquake building height (mu* of about 0.200) and pre-earthquake building height (about 0.195) as the most influential inputs, followed by E-shaped plan configuration and other superstructure materials.

The Sobol results delivered the study’s most striking structural insight. Every first-order Sobol index was close to zero, meaning no single feature independently explains a meaningful share of the model’s output variance. Yet total-order indices told a different story: height_ft_post_eq reached about 0.75, building age about 0.65, and plinth area about 0.60. The enormous gap between first-order and total-order indices indicates that predictions depend heavily on interactions between features rather than on any one attribute acting alone. Building age and floor area, which Morris analysis did not flag, gained importance only in combination with structural features, exactly the kind of conditional behavior single-feature importance measures would miss.

For interpretability, the researchers ran three complementary explainable AI techniques on the trained network: SHAP, a game-theoretic attribution method; LIME, which fits local surrogate models around individual predictions; and ELI5, which uses permutation importance by shuffling features and measuring performance loss. SHAP identified technical_solution_proposed as the most important feature globally, with a mean absolute SHAP value above 0.7, followed by post-earthquake floor count at about 0.42 and post-earthquake height at about 0.25. ELI5’s permutation ranking instead put condition_post_eq first at 0.0360, with mud mortar stone superstructure and pre-earthquake height close behind, an asymmetry suggesting the model relies on post-earthquake observations for class-discriminative predictions while its overall decision structure is most destabilized when pre-earthquake measurements are shuffled.

At the instance level, LIME explained a single building predicted as Grade 4 with full model confidence, showing foundation type, other floor type, and the proposed technical solution as the strongest positive contributors, while post-earthquake condition and mud mortar brick construction pushed against the classification. The consistency of technical_solution_proposed, post-earthquake condition, and post-earthquake floor count across all three XAI methods marks them as the most reliable predictors in the framework, and the authors stress these tools illuminate model behavior rather than optimize it or establish causal relationships.

The authors are candid about the framework’s limits. Many of the top predictive features, including post-earthquake floor counts, heights, and condition assessments, are measurements taken after the event, which may not be available at prediction time in real deployments and could inflate reported accuracy through data leakage. The dataset comes from a single earthquake, limiting generalization to other seismic regions without further validation, and the MLP hyperparameters were chosen from standard configurations rather than systematic optimization. Future work, they write, will focus on pre-earthquake-only models, validation on other earthquake datasets, and expanded sensitivity analysis on other deep learning architectures, along with a user-facing application for researchers, engineers, and disaster management authorities.

Subject of Research: Machine learning prediction of post-earthquake building damage grades using the 2015 Gorkha earthquake dataset

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

Article References: Sharmin, N., Sifat, S. J., & Hossain, M. S. (2026). Predicting earthquake building damage grades using MLP with class imbalance handling global sensitivity analysis and multi method explainability. Discover Civil Engineering, 3(1), Article 258. https://doi.org/10.1007/s44290-026-00650-4

Image Credits: AI Generated

DOI: 10.1007/s44290-026-00650-4

Keywords: earthquake, Gorkha earthquake, machine learning, multilayer perceptron, class imbalance, SMOTE, ADASYN, sensitivity analysis, Sobol indices, explainable AI, SHAP, Nepal