Earthquake Magnitude Prediction Using Machine Learning

A sudden release of energy into the outermost shell of the earth that creates waves of energy that results in shaking of earth's surface is called an earthquake or tremor.

Earthquake prediction is a branch of seismology science concerned with defining the date, location, and magnitude of potential earthquakes within specified limits, and in particular determining the conditions for the next strong earthquake to occur in a area.

Predictions are considered meaningful if progress can be demonstrated beyond random chance. Thus, statistical hypothesis testing approaches are used to assess the probability that an earthquake such as is expected will occur anyway (the null hypothesis). Then, the predictions are tested by checking whether they match better than the null hypothesis with real earthquakes.

An earthquake0precursor is an irregular trend that might give useful warning of an upcoming earthquake. In aeronomy to zoology, there have been 400 reports of possible precursors as defined by science, of 20 different types. But none of them are found useful for earthquake prediction.

Alternative methods for earthquake prediction to be looked for trends or patterns that lead to an earthquake, rather than watching for irregular phenomena that might show precursor that leads to a disastrous earthquake. These patterns are complex and involve many other attributes, advance statistics methods are used to understand the attributes. For earthquake prediction these techniques proved to be more probabilistic and have huge time spans.

Earthquakes being destructive in nature, humans are in dire need of a prediction method. Earthquake prediction means to predict precise time span, intensity and place of a future earthquake. Scientific community has made major contribution but, due to its random phenomena no effective method has been discovered. It is noted that massive earthquakes occur at landscapes where long term notes have been taking place. Some large earthquakes creates a spatial form and certain forecasts regarding value and region are possible. Nevertheless, earthquakes generation isn't always a cyclical process due to the incomplete strain release, the variation of the rupture vicinity and earthquake mediated interactions along different faults. This means that the time among events may be extraordinarily irregular. So, the prediction of the time, or especially near time interval, of an approaching massive earthquake is still the subject of research.

Presently there is no preferred technique for earthquake prediction. Moreover, there may be nevertheless no agreement in science community on whether it is viable to find an answer of this problem. However, rapid improvement of device getting to know techniques and a success software of these strategies to various sorts of problems shows that these technologies could help to extract hidden styles and make correct predictions.

In this paper, machine learning techniques have been used to predict earthquake magnitude for Hindukush area. Statistically calculated eight seismic indicators has been used for earthquake prediction using data set of mentioned regions. The used attributes are based on the well-known geophysical facts of the inverse law of Gutenberg – Richter, the spread of characteristic magnitudes of the earthquake and the seismic quiescence. Machine learning techniques like Pattern Recognition Neural Network, Recurrent Neural Network, Random Forest and Linear Programming Boost Ensemble are used. The model is capable to detect earthquakes of 5.5 magnitude or greater on data of 1 month. Mainly four ML techniques have been used for prediction. Every classifier shows better sensitivity. PRNN shows better results in case of false alarms when compared to other techniques. If geophysical facts are used to design a model to predict non-linear and random phenomena will give more accurate results.

Save your time!
We can take care of your essay

• Proper editing and formatting
• Free revision, title page, and bibliography
• Flexible prices and money-back guarantee

Place Order

[bookmark: _Hlk36672296]Earthquake prediction proves to be a challenging field to conduct research, where a future happening of destructive disaster is predicated. In this study [2], sixty features are calculated by applying seismological concepts, such as seismic rate changes, fore shock frequency, Gutenberg-Richter law, total recurrence time, seismic energy release. Therefore, earthquake forecasting is conducted on the basis of calculated features in place of statistics of earthquakes, therefore by changing a statistical prediction question into a classification question. The computed parameters, are basis to represent the inner geological state of ground before earthquake happens. This study states the mathematical procedures and calculates all features in a attempt to hold most of the data, which gives direction to every earthquake occurrence Et corresponding sixty seismic features. Maximum Relevancy and Minimum Redundancy (mRMR) based parameter selection is applied to select most relevant parameter having maximum information, after getting most of the available seismic attributes. Earthquake prediction model (SVR-HNN) is generated using Support Vector Regression SVR combine with Hybrid Neural Network HNN model and Enhanced Particle Swarm Optimization EPSO, and there state between feature vectors and their concerned Et. The suggested prediction method is applied to execute earthquake predictions separately for the mentioned areas and results are assessed. The proposed prediction model presented promising results when compared to other models suggested for the mentioned regions. In this study, a multistage model is recommended (SVR-HNN). Combination of different machine learning methods are used, with every method 0complementing other via knowledge gathered during learning methods. Therefore, this model is strengthened at every step, gives a more0better final prediction model. Parameter selection is directly affects the performance of SVR, HNN and EPSO. To obtain good results for algorithms SVR, HNN and EPSO, extensive experimentation is carried out to select appropriate parameter values. SVR-HNN show minor deviation, which supports the choice of selected parameters, making a better earthquake prediction model.

In [2], with the computation of maximum seismic features in combination with robust multilayer prediction model is generated. The model constitutes on support vector machine combine with hybrid neural network HNN and Enhanced Particle Swarm Optimization EPSO. For an preliminary assessment SVR is used to predict earthquake, then output is given to HNN as auxiliary predictor in addition with features. HNN is further employed with EPSO. Thus, SVR HNN based model is tested and trained with promising and better results for mentioned regions.to analyze the seismic wave also known as earthquake wave obtained by seismological stations and to ensure the arrival time for S and P wave is the main goal of this paper.AR Pick Algorithm is used to P and S pick times. ASCII format is used to convert waveform in this paper. When time series data was used in time sample sets, results were at all waveform highs and lows with noise.

In [3], a total of 27 features including time-domain features and waveform features were used. BP-NN was outperformed by BP Adaboost and SVM with the accuracy of 100%. Feature extraction is done using SVD. Magnitude prediction and earthquake prediction is done using SVM. The combined accuracy of SVD and SVM methods obtained, was 77% and 66.67% respectively. Machine learning is useful and effective in seismology and reduce mathematical efforts. New features can be added to get improves prediction results that will help in predicting earthquakes. Other environmental features in the active zones can be of good use as a set of attributes in predicting earthquakes and destruction caused by them. Output can be improved by using AR picker algorithm for labelling data. For future work area which can be explored are0statistical and machine learning methods.

In [4], author has discussed about the segregation issue in EEW applying a combination of generative antagonistic systems (GANs) and Random Forests. GANs are solo learning calculations that comprise of two neural systems, a0generator and a pundit, contending with one another. The generator is intended to deliver engineered trials that are as practical as possible so as to trick the analyst, while the analyst is intended to identify the difference between the generator yield and the genuine information. The quake informational collection comprises of broadband and solid movement waveforms from the Southern California Seismic Network (SCSN) and solid movement waveforms from Japan. In the wake of preparing, the discriminator, as a blend of the GAN pundit and the Random Forest classifier, accomplishes 99.2% precision for P waves and 98.4% exactness for clamor flags in the test informational index. As it were, we have 0.8% opportunity to botch a P wave as commotion and 1.6% opportunity to botch a clamor signal as a P wave. Our favored methodology is relevant to numerous other separation as well as recognition errands identified with seismic waveforms. The GAN can undoubtedly distinguish the waveform includes that are generally significant to the objective, which doesn't require abstract decisions of capabilities. The Random Forest fuses extra force by its demonstrated strength in order of various articles. Since these instruments are new to seismology, the uses of their amazing abilities are still under investigation. Consolidating the GAN pundit and the Random Forest, we accomplished the beginning of-the-craftsmanship execution in segregating seismic tremors against other hasty clamor triggers, which can essentially lessen bogus triggers in EEW. Our investigation presents a convincing defense that GANs have ability to find reduced portrayal of seismic waves, which has potential for wide applications in seismology.

In [5], ensemble classification is used with Genetic Programming (GP) which was derived using boosting 0(GP-Adaboost), has been used for the prediction of earthquake (EP-GPBoost). Regions of0Hindukush, Chile and Southern California has been observed for the modelling of indicator using GP AdaBoost. A strong classifier is designed using searching capabilities of GP and boosting of AdaBoost are collectively used to get desired results. These regions have been chosen because a large number earthquake occurred in these regions. In this study, EPS is modelled as a binary classification problem with the goal of generating predictions for 5.0 and greater magnitude earthquakes, 15 days before an earthquake. The EP-GP0Boost shows remarkable performance for all three regions, particularly in terms of low false alarms generation. The accuracy of 74%, 80% and 84% for Hindukush, Chile and Southern California means that the false alarm ratio is significantly low. Proposed model shows improved results for these regions when compared with other studies. Future research will concentrate on the development of more effective seismic indicators and the implementation of deep learning techniques for earthquake prediction.

In [6], one of the best predictive method, support vector regression (SVR) has been used. The dataset and SVR training parameters 0influence 0the overall performance of the model. In this paper, particle filter method is used to improve the results of SVR. The particle filter, which is typically set by try and error, defines the SVR parameters in the proposed process. The data used in this analysis were collected from two Iranian databases, climate data from the Meteorology Center of the Islamic Republic of Iran and seismic data from the Iranian Seismology Center. The model could show predictive accuracy and the number of earthquakes expected in a month. By 96 % accuracy for mean magnitude and over 78 % for the number of earthquakes. If the proposed method has been able to classify results within the daily range and distinctly different results can be identified using precursor forms, since the data provided was restricted to monthly data.

In this paper [7], authors have A type of machine learning which can predict earthquake with improved results. Authors used historical earthquake data obtained from the Bangladesh Meteorological Department (BMD) to conduct this experiment. Experiment cannot be done using raw data directly. Preprocessing has been done on the data to perform experiment. Proper preprocessing and optimizing has been done in different steps to prepare experimental dataset.0The training dataset contains 80% data and training dataset contains 20% data. The ratio of training and the testing dataset was predefined with a specific cross-validation method. Support Vector Machine (SVM), Random Forest, Naïve bays, Decision tree, and K-nearest neighbors were applied for training the model. The model parameters were also precisely fitted to the cross-validation process to obtain precise results. The0results of the training0were0compared and0evaluated0among the0models to fine-tune the input variables and model0parameters to increase the0predictability. The intensity level in terms of prediction of earthquake was slightly better than others in case of Random forest. The Support Vector Machine model with a linear kernel, however, also predicted good results which were almost identical to the random forest model. This paper's section on the results analysis shows that the accuracy is very good in predicting earthquake phenomena, but it is not that simple. This model only conducts a statistical study of historical data concerning the earthquake. However, far more detailed data such as geological data, geographical data, tectonic plates moving data are required to get the most precise and consistent observations about earthquake prediction.

In paper [8]., author has presented signal detection and classification of high noise level and low correlation ground motion signals, and proposed a recognition and a classification technique based on support vector machine (SVM) model enhanced by genetic algorithm. Classifiers are designed by recognition of ground targets and ground-motion signal processing, ground-motion signal analysis and feature extraction has been mainly used in this process. Eigenvectors can train and test on improved version of SVM based on genetic algorithm. The accuracy of test is 93.75%. Recognition efficiency and higher training speed and performance of classifier can be improved. Adaptive denoising algorithm can be used for signal preprocessing, and recognition technology incorporating multiple classification methods can be used to optimization. Support Vector Machine and Neural Network combination, with fuzzy theory and so on to improve target classification efficiency and accuracy, are areas in which more work can be conducted to establish a more efficient target recognition method for ground movement signals.