Thermal anomaly detection using NARX neural network method to estimate the earthquake occurrence time

The researchers’ view toward earthquake has been vastly changed from the past until now in such a way that the field of earthquake prediction and assessing its pre-indicators have been receiving significant attention from researchers. By the development of remote sensing techniques and obtaining thermal data of the earth’s surface and different layers above it, it has become possible to properly study the thermal changes before, during, and after the earthquake. In this study, the remotely sensed thermal data from the earth’s surface of the center zone of the earthquake was used in order to predict the time of earthquake occurrence. To date, smart methods which include Artificial Neural Networks (ANN), Support Vector Machine (SVM), Genetic Algorithm (GA), etc., contain different uncertainties depending on their training algorithms in such a way that by defining an inappropriate threshold between the predicted value and the real ones, they are not able to isolate the variable but natural behavior of the under-study area from the anomaly. For instance, they identify the natural increase in the environment temperature, which could be as a result of seasonal or climatic conditions, as a thermal abnormality which could enter high levels of errors in determining the time of earthquake occurrence. Considering the fact that a series of time-dependent data should be used in studying earthquakes, the prediction of these time series can be done using Artificial Neural Networks. In order to make it more accurate, two different methods of dynamic NARX (Nonlinear Auto Regressive with eXternal input) neural network algorithm namely Levenberg-Marquardt and Scaled conjugated gradient have been applied. After that, the responses of these two methods have been compared with the response derived from mean and variance. The important advantage of the NARX neural network is that it can detect and consider small thermal anomalies caused by natural climate change, which cannot be done by regular earthquake pre-indicators. The results elucidate that the Saravan earthquake, 5 days before (Levenberg-Marquardt method response) and 11 days before (Scaled conjugated gradient method response), Goharan earthquake 13 days before and Borujerd earthquake 8 and 6 days before occur has been predicted. The thermal anomaly about 5, 7 and 7 degrees of Kelvin of earthquakes center zone respectively detected. Thus the thermal anomaly detected by this method can be a good pre-indicator for earthquake prediction..