Detection and Classification of Sporadic E Using Convolutional Neural Networks

In this work, convolutional neural networks (CNN) are developed to detect and characterize sporadic E (Es), demonstrating an improvement over current methods. This includes a binary classification model to determine if Es is present, followed by a regression model to estimate the Es ordinary mode critical frequency (foEs), a proxy for the intensity, along with the height at which the Es layer occurs (hEs). Signal-to-noise ratio (SNR) and excess phase profiles from six Global Navigation Satellite System (GNSS) radio occultation (RO) missions during the years 2008-2022 are used as the inputs of the model. Intensity (foEs) and the height (hEs) values are obtained from the global network of ground-based Digisonde ionosondes and are used as the "ground truth," or target variables, during training. After corresponding the two data sets, a total of 36,521 samples are available for training and testing the models. The foEs CNN binary classification model achieved an accuracy of 74% and F1-score of 0.70. Mean absolute errors (MAE) of 0.63 MHz and 5.81 km along with root-mean squared errors (RMSE) of 0.95 MHz and 7.89 km were attained for estimating foEs and hEs, respectively, when it was known that Es was present. When combining the classification and regression models together for use in practical applications where it is unknown if Es is present, an foEs MAE and RMSE of 0.97 and 1.65 MHz, respectively, were realized. We implemented three other techniques for sporadic E characterization, and found that the CNN model appears to perform better. Ionospheric Sporadic E (Es) are cloud-like structures of dense ionization in the Earth's upper atmosphere. As radio waves from Global Navigation Satellite System (GNSS) satellites propagate through these layers of irregular plasma, phase and amplitude perturbations may be introduced into the signals. GNSS radio occultation (RO) missions receive these perturbed signals and can infer Es intensity and height characteristics on a global scale. As GNSS-RO missions do not directly measure foEs and hEs values, ground-based ionosondes can be used to provide true values on which to train and validate models. In this work, data from several GNSS-RO missions and ionosondes between 2008 and 2022 were used. While previous approaches have used more traditional signal processing methods, here we use machine learning methods to develop the models. These models are trained by ingesting the GNSS-RO data and learning the best estimating function that minimizes the error between predicted values and the true values provided by the ionosondes. To ensure both the GNSS-RO and ionosondes are measuring the same physical phenomena, we use a window of 150 km and 30 min to join the data. The models trained using machine learning methods demonstrate improved performance when compared with other methods described in literature. CNN models were developed to detect and characterize sporadic E layers using radio occultation SNR and excess phase profilesModels explored using both the ionosonde foEs values and intensity focused on Es metal ion layers, fo mu EsMachine learning models demonstrate the ability to skillfully extract Es parameters from radio occultation measurements