Vs30 engineering geologic interpolation: synthetic augmentation and machine learning in data-deficient environments

The paper examines synthetic augmentation of spatial data using the Residual Bootstrap method and a hybrid KMeans–SMOTENC scheme to improve the interpolation quality of the Vs30 parameter when field measurements are limited. The principles of both approaches are described: RF-bootstrap preserves the structure of the original data by adding empirical residuals to predictions based on a random forest, while the hybrid scheme combines global KMeans clustering with local SMOTENC interpolation. A comparative analysis of their interaction with machine-learning algorithms across different modeling scenarios and varying point densities shows that for small samples (fewer than 15 points), the Residual Bootstrap approach is more effective, yielding the lowest MAE and RMSE errors, whereas for more representative datasets (≥25 points), the KMeans–SMOTENC hybrid performs better due to improved reproduction of local structure. An adaptive strategy is proposed for selecting the augmentation method depending on the size and homogeneity of the initial sample, enabling a substantial increase in the accuracy and robustness of machine-learning-based Vs30 interpolation under data-scarce conditions.

Ҳали таржима қилинмаган