Modeling ionograms with Deep Neural Networks: Applications to Nowcasting

The state parameters of the ionosphere are of fundamental importance not only for space weather studies but also for technological applications such as satellite radio communications. As with many geophysical phenomena, the ionosphere dynamics are governed by nonlinear processes that make ionospheric forecasting a challenging endeavor. However, we have enormous datasets and ubiquitous experimental sources that can help us find the complex regularities in these phenomena. We forecasted ionograms for different solar activity times and database sizes using regression deep neural networks. Due to the neural network's extrapolation of virtual heights for all frequencies given to the model, we estimated foF2 using two embedded different models to identify the last frequency of each ionogram. Furthermore, we made hyperparameter tuning for each training set applying the k-fold cross-validation method. The predictions were compared to measurements collected with the Digisonde system at the Jicamarca Radio Observatory, a persistence model, IRI, and the SAMI2 model estimations. Finally, we will present preliminary results on a new virtual heights model that predicts the difference between consecutive ionograms and preliminary results from a model to estimate electron densities.