A back-propagation artificial neural network model on microtremor data to evaluate liquefaction (Case Study: Yogyakarta City, Indonesia)

Abstract

Yogyakarta City lies in a seismically active graben basin filled with young, unconsolidated volcanic deposits and is susceptible to earthquake-induced liquefaction. Conventional screening reduces each Horizontal-to-Vertical Spectral Ratio (HVSR) curve to a single resonance parameter, which cannot represent the multiple impedance contrasts of a heterogeneous volcanic basin. This study tests whether descriptors of curve morphology can reproduce a liquefaction susceptibility classification, using a Back-Propagation Artificial Neural Network (BP-ANN). The analysis uses 110 ambient-noise records processed with BIDO 2.0 and screened against SESAME reliability and clear-peak criteria. Each curve was encoded as an eleven-element peak and valley vector, of which the archived parameters populate four. Target classes were derived from Liquefaction Potential Index values at 27 boreholes, ten logged for this study, and 29 of 110 stations are susceptible. Hidden-layer width was selected by systematic search, the loss weighted inversely to class frequency, and performance estimated by stratified five-fold cross-validation repeated ten times. The spectral descriptors give an accuracy of 53.3 ± 10.4 %, below the 73.6 % obtained by always predicting the majority class, and an F1-score of 35.8 ± 14.9 %. Widening the hidden layer fivefold changes validation loss by less than 0.01, so the limitation lies in the input, not model capacity. The same network given shear-wave velocity and groundwater depth, on which the classes depend, reaches 88.5 ± 5.9 %: the classification is recoverable from these parameters but not from HVSR peak frequencies, and single-split accuracies should be read with caution.