Forecasting the Magnitude Category Based on The Flores Sea Earthquake

Abstract

Earthquakes are a phenomenon that is still a mystery in terms of predicting events, one of which is the magnitude. As technology develops, there are many algorithms that can be used as approaches in earthquake forecasting. In the context of magnitude forecasting, the application of GaussianNB, Random Forest and SVM has the potential to reveal these patterns and relationships in the data. With the six main phases of this research, namely data acquisition, data preprocessing, feature selection, model training, forecasting result evaluation, and performance analysis. From these results we obtain, firstly that the GaussianNB model has a relatively simple and fast method in training its model. However, the weakness lies in the assumption of a Gaussian distribution which may not always suit the complex and diverse characteristics of earthquake data. Based on GaussianNB model, the model accurately predicts magnitude category 1 for 421 observations and magnitude category 2 for 33 observations. Meanwhile, the magnitude 3 and magnitude 4 categories did not produce accurate predictions from the model. Second, Random Forest, this method can increase accuracy and overcome the overfitting problem that occurs when forecasting magnitudes. In contrast to GaussianNB, it tends to result in models with greater complexity and require more time to compute. In our findings, we obtained an MSE value of 0.12 with an R2 score of-0.10, this indicates conditions that are less effective in explaining differences in test data. The third option is SVM, which has both benefits and drawbacks that must be taken into account. The capacity of SVM to separate data that has both linear and non-linear separation is one of its key advantages; nevertheless, the main drawback is that it is sensitive to hyperparameter adjustments. It is clear from the results of the algorithm comparison that SVM has more potential for earthquake forecasting, especially the linear SVM and polynomial SVM model. The accuracy of the standard SVM is 0.587, which indicates relatively low performance. Linear SVM obtained a very high accuracy of 0.998. Meanwhile, Polynomial SVM achieves perfect accuracy of 1.0. while RBF SVM has the same accuracy as standard SVM, namely 0.587.