Random forest prediction of ACeBS seismic resilience scores for simple residential buildings in Yogyakarta

The Asesmen Cepat Bangunan Sederhana (ACeBS) is a 47-item checklist used in Indonesia to rate the earthquake resilience of simple residential buildings, but its scoring still relies on manual field tabulation by trained surveyors, which is slow and difficult to scale. This study develops and validates a machine-learning model that predicts the ACeBS resilience score directly from the recorded checklist answers and building age, and identifies which assessment items most influence the score. A dataset of 110 single-storey houses surveyed across the five regencies of the Special Region of Yogyakarta was used; each record contains binary responses to the 47 ACeBS items, the building age, and the percentage resilience score. A Random Forest regressor with 100 trees was trained on 100 buildings using ten-fold cross-validation and tested on 10 unseen buildings, with R-squared, mean absolute percentage error (MAPE) and root-mean-square error (RMSE) as performance metrics and Gini importance for parameter screening. The model reproduced the resilience score with high accuracy, reaching a cross-validation R-squared of 0.853 (MAPE 14.6%, RMSE 8.56) and an evaluation R-squared of 0.922 (MAPE 5.31%, RMSE 4.21). Feature-importance analysis showed that three checklist items (P21, P20, P12) together with building age accounted for the majority of the predictive power, and that 32 of the 48 inputs exceeded the 0.005 retention threshold. The findings indicate that ACeBS scoring can be reliably automated and reduced to a smaller set of decisive items, supporting faster and more consistent pre-disaster screening.