A Deep Neural Network Model of Global Topside Electron Temperature Using Incoherent Scatter Radars and Its Application to GNSS Radio Occultation

The goal of this study is to present a new model for global topside electron temperature (Te) using a deep neural network (DNN) that is trained using measurements from Incoherent Scatter Radars (ISRs). This study is also an investigation into whether this model can be used to generate the electron temperature in the topside ionosphere using GNSS ionospheric radio occultation (GNSS-IRO) data as the input. 

This study uses a DNN technique to create a new global topside electron temperature model from three sub-models that have been trained using data from three ISR stations; Arecibo (low-latitude), Millstone Hill (mid-latitude) and Poker Flat (high-latitude). This global model is trained using electron density profile information (e.g., vertical scale height, hmF2 and NmF2) and solar and geomagnetic activity (F10.7 and Kp, respectively) in addition to traditional spatial and temporal variables (e.g., local time, month and latitude) as the independent variables. After the Te model is developed, Te information can be generated from the GNSS-IRO electron density profiles using a newly created Ne-Te model. This model's outputs are assessed with regards to out-of-sample ISR data and compared to the latest International Reference Ionosphere (IRI) model. It is found that the electron temperature profiles from the DNN have an RMSD of 259K in the low-latitude region (i.e., against Arecibo data), 254K in the mid-latitude region (i.e., against Millstone Hill data) and 314K in high-latitude region (i.e., against Poker Flat data), and all of them are smaller than the RMSD from IRI. An additional comparison between the model results versus the Thermosphere-Ionosphere-Electrodynamics General Circulation Model (TIEGCM) outputs is also conducted. A statistical analysis of the diurnal electron temperature profiles obtained from GNSS-IRO is shown to agree with the TIEGCM outputs.