Machine Learning for Estimation of Surface Ground Structure by H/V Spectral Ratio

A new method for efficiently identifying ground motion observation site is proposed by the deep learning with a convolutional neural network (CNN). This method transforms the acceleration spectra at an observation site into unique color spectra corresponding to their amplitudes, which are then learned by the deep learning. First, the seismic H/V spectral ratios of earthquake motions of 50 gal or less obtained at eight K-NET stations are tested. As a result, the K-NET sites were identified with an accuracy of more than 95%. Misclassified earthquake motions were unique ones such as those having large source distances or hypocenter depth being deep. Next, the microtremor H/V spectral ratios obtained at the same K-NET sites were input to the CNN which was solely trained by the seismic H/V spectral ratio as above. The accuracy varied from 0 to 95% and on average 50%. Although results are encouraging, the method requites further developments.