A Critical Investigation on the Reliability of GPS-Derived TEC Data at Agra for Earthquake Predictions by Using the Support Vector Machine (SVM) Algorithm

Abstract

In the present paper, a long period of GPS-TEC observed at Agra station, India has been investigated to develop a support vector machine (SVM)-based model corresponding to earthquakes that occurred around Agra within 2000 km from 2010 to 2013. The different datasets are prepared with the help of the GPS‑TEC data, solar activity (R and F10.7 cm), magnetic storm (Dst and ∑Kp indices) parameters, etc. for the magnitude in the range of 4–7.7 at the interval of 0.5 and outer radius of 2000 km at the interval of 500 km. Here, 90% of the data is used for training, and the rest of the 10% of data is used for test purposes. These parameters are used as the input in the Linear and Medium Gaussian SVM models, respectively. The confusion matrices are obtained for each dataset and then skill scores such as precision, recall, and accuracy are calculated, statistically. The best results of skill scores are obtained in the case of magnitude range 4–7.7 and the outer radius of 2000 km. The receiver output characteristics (ROC) curves are plotted and the maximum accuracy of 94% is obtained. A k-fold cross-validation (k = 5) technique is used to validate our models. Further, both models are compared by using the Wilcoxon signed-rank test. Bayesian Optimizer optimizes the GPS-TEC datasets. Finally, these models may be utilized for real-time data classification.