Reconstruction of Incoherent Scatter Radar Vertical Electron Density Profiles Using Regression Trees

The key point to understanding the ionospheric variability caused by solar and geomagnetic space weather is the ability to rapidly determine accurate electron density profiles wherever necessary across the globe. However, developing a high-predictability model that accurately predicts electron density is still challenging. This study proposes a technique to reliably reconstruct ionospheric vertical electron density profiles for corrupted and missing incoherent scatter radar (ISR) data using a machine-learning-based technique, particularly regression trees (RT). The technique is built using an extensive dataset that includes observations from the Jicamarca ISR ranging from 1997 to 2022, collocated GNSS radio occultation data during 2007-2020, and collocated ground-based vertical incidence sounding (ionosonde) data. The data was split randomly for training and validation using a 10-fold cross-validation technique. To include the solar and geomagnetic activity information, we have used the F10.7 (solar radio noise flux at 10.7 cm wavelength) index and the disturbance storm time (Dst) index. Only data from geomagnetically quiet times are considered in this work. Fine regression tree (RT), coarse regression tree, boosted RT, and bagged RT are the four models considered in this work. The bagged RT is the proposed machine learning model best suited for our technique. The results show that the minimum root mean square error (RMSE) was obtained when using bagged RT with a value of 0.95815 × 10¹¹ and the coefficient of determination (R-squared) was 0.95. However, after applying Bayesian optimization on the bagged RT, the R-squared value increased to 0.70. These initial findings indicate the ability of the proposed model to reconstruct ionospheric vertical electron density profiles in a considerably accurate manner. However, further optimizations and enhancements in the hyperparameters selection are being done. Therefore, the results of our model are expected to improve significantly.