A comparison of classical and intelligent methods to detect potential thermal anomalies before the 11 August 2012 Varzeghan, Iran, earthquake (&amp;lt;i&amp;gt;M&amp;lt;/i&amp;gt;&amp;lt;sub&amp;gt;w&amp;lt;/sub&amp;gt; = 6.4)

M Akhoondzadeh

doi:10.5194/nhess-13-1077-2013

Abstract

Abstract. In this paper, a number of classical and intelligent methods, including interquartile, autoregressive integrated moving average (ARIMA), artificial neural network (ANN) and support vector machine (SVM), have been proposed to quantify potential thermal anomalies around the time of the 11 August 2012 Varzeghan, Iran, earthquake (Mw = 6.4). The duration of the data set, which is comprised of Aqua-MODIS land surface temperature (LST) night-time snapshot images, is 62 days. In order to quantify variations of LST data obtained from satellite images, the air temperature (AT) data derived from the meteorological station close to the earthquake epicenter has been taken into account. For the models examined here, results indicate the following: (i) ARIMA models, which are the most widely used in the time series community for short-term forecasting, are quickly and easily implemented, and can efficiently act through linear solutions. (ii) A multilayer perceptron (MLP) feed-forward neural network can be a suitable non-parametric method to detect the anomalous changes of a non-linear time series such as variations of LST. (iii) Since SVMs are often used due to their many advantages for classification and regression tasks, it can be shown that, if the difference between the predicted value using the SVM method and the observed value exceeds the pre-defined threshold value, then the observed value could be regarded as an anomaly. (iv) ANN and SVM methods could be powerful tools in modeling complex phenomena such as earthquake precursor time series where we may not know what the underlying data generating process is. There is good agreement in the results obtained from the different methods for quantifying potential anomalies in a given LST time series. This paper indicates that the detection of the potential thermal anomalies derive credibility from the overall efficiencies and potentialities of the four integrated methods.

Highlights

11 August 2012 Varzeghan, Iran, earthquake (Mw = 6.4)
In order to quantify variations of land surface temperature (LST) data obtained from satellite images, the air temperature (AT) data Land surface temperature (GLSeTo) sacnoiemnaltiiefsicmay arise due derived from the meteorological station close to the earthquake epicenter has been taken into account
A pre-seismic thermal anomaly is an abnormal increase ily implemented, and can efficiently act through linear solu- in LST that may be observed around 1–10 days prior to an tions. (ii) A multilayer perceptron (MLP) feed-forward neural network can be a suitable non-parametric method to detect the anomalous changes of a non-linear time series such earthquake, with increasesGofeteomspceriaetunretiofincthe order of 3

Summary

Methodology

Variations of the land surface temperature depend on season, geographic location, climatological conditions and other unknown parameters. The training set is used for construction of the neural network, whereas the test set is used for measuring the predictive error of the model. In order to determine the best network configuration, the effective parameters, which influence the value of predictive error, including the number of pattern input, lag value: the number of hidden layers and their number of neurons, the activation functions and the learning algorithm, have been obtained via an iterative process to assess the minimum predictive error when the training process was implemented. At each step, using the training data, the SVM method is implemented and the predictive error (Eq 2) is minimized during the validation of data, where Xi and Xi in Eq (2) are the observed value and the output from the SVM method, respectively. In the case of the testing process, if the value of DXi (i.e. the difference between the actual value Xi and the predicted value Xi) is outside the pre-defined bounds, M ± I QR, the anomaly is detected

Implementation

Findings

Discussion and conclusions