Akademik Veri Yönetim Sistemi
IEEE Access, cilt.13, ss.116200-116210, 2025 (SCI-Expanded)
This study proposes a novel deep learning approach for predicting significant earthquakes (Mw ≥ 5.0) in Turkey using ionospheric Total Electron Content (TEC) data and space weather indices. ConvMixer is a lightweight CNN architecture that blends spatial and channel information using depthwise convolutions and pointwise layers. Inspired by vision transformers, it offers efficient image classification with fewer parameters and high performance. We developed a multi-input one-dimensional convolutional mixer (MI-1D-ConvMixer) model to classify TEC data from the preceding five consecutive days as either precursory to an earthquake on the 6th day or normal. The model incorporates six inputs: five 1D TEC signals and one 1D space-weather index array, including the global geomagnetic index (Kp), storm duration distribution (Dst), sunspot number (R), geomagnetic storm index (Ap-index), solar wind speed (Vsw), and solar activity index (F10.7) are also utilized to reveal non-seismic related pre-earthquake ionospheric variations. Our methodology involves two stages: (1) a preprocessing stage to enhance TEC signals, and (2) an end-to-end training of the MI-1D-ConvMixer model. The model architecture features depth-wise and point-wise convolutions with patch embedding, utilizing four tunable variables: network depth, hidden dimension size, kernel size, and patch size. We used 196 earthquakes data from Turkey from 2010-2023, and TEC data from the TNPGN-Active GNSS stations. The dataset was split into 75% for training and 25% for testing. Performance metrics, including classification accuracy, sensitivity, specificity, and F1-score, are used for evaluation. Our model achieved a classification accuracy of 97.49%, demonstrating its potential for earthquake prediction systems. This research contributes to the field by introducing a novel deep learning architecture specifically designed for integrating TEC and space weather data for earthquake prediction. Future work should focus on validating the model’s performance in different geographical regions and investigating its limitations.