A machine learning algorithm with random forest for recognizing hidden control factors from seismic fault distribution

초록

Stress interaction among many faults in earthquake prone region is strongly affected by the spatial distribution of faults. To build a robust Machine Learning (ML) algorithm for finding hidden relationships between fault distribution and its controlling factors, the performance of ML in dependence of various parameters, such as the complexity and data size of the fault distribution, needs to be investigated. We have, therefore, developed an ML algorithm by combining Principal Component Analysis (PCA) with a Random Forest (RF) to unveil the controlling factors on seismic fault distribution. We synthesized fault images for training of RF classifier N-D (> 10,000) fault images, which scattered the faults in a square with R km R km. R, ranging from 200 km to 600 km, is a controlling factor of fault distribution. PCA extracts the dominant features of the seismic fault images to supply the refined training data to the RF. This leads to a dimensional reduction of the data. We tested the influence of the number of principal components (N-P) on PCA performance in retaining the information. Larger and more variable R values (i.e., a more complicated fault distribution) with larger amounts of training data required more principal components for a full preservation of the original data. We quantified the optimized N-P/N-D ratio for an efficient RF classifier, which was about similar to 0.1 for the evaluating efficacy of the PCA. This showed that the RF classifiers is effective in predicting R values depending on the fault zone with R km R km and not on specific fault locations. We propose that a small N-P, equivalent to similar to 10% of the training data, suffices for training accurately RF classifier for fault distributions, which can reduce the complexity of fault distribution to understand stress interaction among many faults.

키워드