Predicting earthquake magnitude error with regression, decision tree & random forest

Abstract

As important as it is challenging, earthquake prediction plays an integral part in minimizing the catastrophic impact from earthquakes. In Japan, earthquake prediction leads the way by utilizing cutting-edge prediction technology that is applied to the Earthquake Early Warning (EEW) system to provide enough time for people to evacuate to a safe place. This early warning system has achieved a hit rate of 56% as of 2011, with 44% of the time earthquakes can be underestimated or overestimated by a large degree. Such inaccuracy can create panic in the country and cause unease or disastrous effect for the country in the long term. As such, studying the magnitude error can be a way to minimizing the error and reducing inaccuracy. Until recently, there has yet been a study conducted on magnitude error by the machine learning methods. Therefore, this research aims to reduce the magnitude error by predicting the possible error. These are achieved by utilizing the regression, decision tree and random forest analysis. These models are constructed based on the available parameters (magnitude, depth, number of stations, etc.). Subsequently, the result showed a high correlation between magnitude error and the number of recording stations available, which intuitively can reduce the magnitude error value through increasing the number of recording stations. Through the evaluation using the statistical features of the models, two optimal models are selected as the main results of this research.