Modelıng Of The Ionosphere's Dısturbance Usıng Deep Learnıng Technıques
Abrı Zangabad, Rahem
xmlui.mirage2.itemSummaryView.MetaDataShow full item record
The ionosphere drives an essential role in the atmosphere and earth. Solar flares due to coronal mass ejection, seismic movements, and geomagnetic activity cause deviations in the ionosphere. The main parameter for investigating the structure of the ionosphere is Total Electron Content (TEC). This thesis converges the importance of ionosphere TEC data to evaluate seismic events. The dataset assessed in this thesis contains the ionospheric variability during moderate and severe earthquake events of varying strengths for 2012-2019 years in Chile station. TEC values obtained from GPS stations provide a powerful technique for analyzing the ionospheric response to earthquakes and solar storms. TEC data gathered from GPS stations (Dual-Frequency GPS receiver) is used to investigate the ionospheric variability through moderate and severe earthquakes. This thesis has three main contributions. In the first contribution, our goal is to analyze the relations between earthquakes and TEC data. We concentrate on extracting features from earthquakes and classification over the ionospheric TEC data. In this phase, we do not focus on predicting earthquakes with previous days. The proposed model uses Deep Autoencoders to extract features from TEC data. As the ionospheric TEC data is a high-dimensional factor, reducing dimensionality to obtain a compressed feature set is an essential step in the feature extraction phase. The collected features served as input to dense neural networks to perform classification. The classification model results are compared against the LDA(Linear Discriminant Analysis), SVM(Support Vector Machine) and Random Forest classifier models to evaluate the proposed model. The results report that the proposed model improves in distinguishing the earthquakes at an accuracy rate of about 0.94 in the target station zone. In the second contribution, we propose a classification model to detect earthquakes in previous days. The LSTM methods handle this issue with the solution to short-term memory. The proposed models use LSTM-based (Long Short-Term Memory), deep learning models to classify earthquakes days by analyzing TEC values of the last seven days. The variant versions of the LSTM models are proposed to enhance the contribution of this research. The LSTM-Based classification models are compared against the SVM, LDA and Random Forest classifier models to evaluate the proposed models. The results reveal that the proposed models improve in detecting the earthquakes at an accuracy rate of about 78-80 and can be used as a successful tool for detecting earthquakes based on the previous days. In the last contribution of this thesis, we develop a hybrid version of deep autoencoders and LSTM to detect earthquakes in previous days. This model proposes to improve the stacked LSTM-based earthquake classification introduced in the second model. The suggested model uses a deep autoencoder to derive beneficial features from ionospheric TEC data and perform Stacked LSTM to classify earthquakes days by analyzing TEC values of the last seven days. To analyze the contribution of the suggested DAE-STCK-LSTM model, we used Stacked-LSTM, LDA, and SVM classifiers. Our evaluation test results prove approximately 81-84 accuracy-based performance in the two test sets of the earthquakes, including moderate and severe earthquakes.
The following license files are associated with this item: