Comparative Study of Various Pattern Recognition Techniques for Identifying Seismo-Tectonically Susceptible Areas

Abstract

One of the most destructive natural hazard is an earthquake. Both the built environment and significant lives are lost. Due to presence of subduction zone and continuous continent to continent collision, major earthquakes in India are concentrated along the seismo-tectonically active Himalayan arc. Seismicity along the Alpine Himalayan Belt varies in well-defined patterns which are based on different parameters such as distance from tectonic source, presence of other historic earthquakes, intersections of tectonic features (in numbers), river length, presence of water body, meandering of rivers, soil conditions. An earthquake catalogue consisting of 540 earthquakes from year 1552 to 2020 is prepared for the consideration of data set. 117 tectonic features are considered for the study. For extracting feature a buffer of radius 25 km is considered along each earthquake of the catalogue. Four features namely distance from nearest tectonic feature, magnitude of earthquake, number of earthquakes within the buffer zone and presence of rivers (length) is extracted. A combination of these features allows identifying patterns based on training and classification of the entire data set.

This paper is an attempt to compare four classification algorithms namely Linear Discriminant Analysis (LDA), Support Vector Machines (SVM), K nearest neighbors (KNN) and Artificial Neural Networks (ANN) and present an integrated predictive model for identifying seismically susceptible zones (SSZ) around the Kangra Region. The seismicity in the study area is divided into two classes: class A and Class B based on the magnitude of earthquakes and are called as training data set for the classification algorithms. The training data set provides a basis for classifying a set of data into class A and Class B based on the extracted features. After classification the study area is the classified into two classes: Class A: Significant seismically susceptible area, Class B: less Susceptible zone. A comparative analysis is done in terms of accuracy of classifying an earthquake into the correct class as per training data set. Of the four classification methodologies, ANN showed the best performance by reducing the error in misclassifying the epicenters from either class. The errors accuracy of classifying an earthquake in the correct group were found to be 24.74%, 6.45%, 7.78%, 2.9% for LDA, SVM, KNN and ANN respectively.