A New Feature for Prediction of Significant Earthquakes by Artificial Neural Networks

Earthquakes happen continuously, but those that are of higher magnitude are more important to detect, because their detection can be directly used for early warning, evacuation and preventing loss of life. However, at present there is no scientific method that can predict earthquakes a reasonable amount of time (at least a few days) in advance. Earthquakes are very unpredictable, and their nucleation process is not completely understood, and local parameters play an important role, making it difficult to build a generalized model that applies to all earthquakes in different fault lines in different locations. In this study, I have applied machine learning techniques to fundamental parameters from data published by Kandilli Earthquake Research Center and US Geological Survey, as well as derived parameters proposed by other researchers and compared the results to those calculated with a new spatial feature that considers the spatial clustering of earthquakes in three dimensions before a major shock. The results show that the spatial feature may improve accuracy in some cases and demonstrate how accuracy is improved by the new feature.