An investigation of new ionospheric models using multi-source measurements and neural networks

Hu, A 2019, An investigation of new ionospheric models using multi-source measurements and neural networks, Doctor of Philosophy (PhD), Science, RMIT University.


Document type: Thesis
Collection: Theses

Attached Files
Name Description MIMEType Size
Hu.pdf Thesis application/pdf 840.07KB
Title An investigation of new ionospheric models using multi-source measurements and neural networks
Author(s) Hu, A
Year 2019
Abstract Ionosphere is one of the atmospheric layers that has a major impact on human beings since it significantly affects the radio propagation on Earth, and between satellites and Earth (e.g., Global Navigation Satellite Systems (GNSS) signal transmission). The variation of the electrons in the ionosphere is strongly influenced by the space weather due to solar and cosmic radiation. Hence, the short/long-term trend of the free electrons in the ionosphere has been regarded as very important information for both space weather and GNSS positioning.

On the other hand, precisely quantifying the distribution and variation of free electrons at a high spatio-temporal resolution is often a challenge if the number of the electrons (electron density) is detected only from the traditional ionospheric sensors (e.g., ionosonde and topside sounder and Incoherent Scatter Radar (ISR)) due to their low spatio-temporal coverage. This disadvantage is also inherited from the empirical ionospheric model developed based on these data sources. Nowadays, the availability of advanced observation techniques, such as GNSS Radio Occultation (RO) and satellite altimetry, for the measurement of Electron Density (Ne) and related parameters (e.g., hmF2, NmF2, Vertical Scale Height (VSH), Electron Density Profile (EDP) and Vertical Total Electron Content (VTEC)) in the ionosphere has heralded a new era for space weather research in the upper atmosphere. The new sources of data for ionospheric modelling can improve not only the accuracy but also the reliability of the model (such as[96] for hmF2 and [28] for VTEC).

In this study, Helmert Variance Component Estimation (VCE) aided Weight Total Least Squares (WTLS) is selected for modelling global VTEC using International GNSS Service stations, satellite altimetry and GNSS-RO measurements. The results show that the new VTEC model outperforms the traditional global ionospheric VTEC Model by at least 1.5 Total Electron Content Unit (TECU) over the ocean. This improvement is expected to be significant in the refinement of global ionospheric VTEC Model development. As is well known, the most traditional models developed are prone to the effects of inherent assumptions (e.g. for the construction of the base functions in the models) which may lead to large biases in the prediction. In this study, an innovative machine learning technique (i.e. Neural Network (NN)) is investigated as the modelling method to address this issue. Different from the traditional modelling method, neither the observation equations (or the so called `design matrix'), nor apriori knowledge of the relationship (both of them can be considered as the source of the aforementioned assumptions) is required in the modelling process of a NN. This network system can automatically construct an optimal regression function based on a large amount of sample data and the designed network [43].

In this study, Deep Neural Network (DNN), which is an advanced Artificial Neural Network (ANN) (with more than one hidden layer), is investigated for their usability of VSH and topside EDP modelling, as well as the relationship between Ne and electron temperature. The results reveal that the new VSH model agrees better than the traditional model with regards to either out-of-sample measurements or the external reference (i.e. ISR data). In addition, the new model can represent the characteristic of VSH in the equatorial region better than that of traditional approaches during geomagnetic storms. The relationship between Ne and Electron Temperature (Te) investigated from ISR data can be used to improve the performance of the current Te model. The local time-altitude variation of the model outputs agrees well with that from a physical model (i.e., Thermosphere-Ionosphere-Electrodynamics General Circulation Model (TIEGCM)). The new topside EDP model takes hmF2 and NmF2 into consideration as part of the variable set. Comparing with the reference data (i.e., out- of-sample COSMIC data, GRACE and ISR data), the new model agrees much better than the International Reference Ionosphere (IRI)-2016 model. In addition, an advanced NN technique, Bidirectional Long Short-Term Memory (Bi-LSTM), is utilised to forecast hmF2 by using the hmF2 measured by Australian ionosondes in the five hours prior. The forecast results are better than the results from real-time models in the next five hours. The new model performs also better than the current hmF2 model (i.e., AMTB [2] and shubin [96] models, which is used inside IRI-2016 model) by at least 10km in most ionosonde stations.

Overall, the neural network technique has a great potential in being utilised in the ionospheric modelling. In addition to the accuracy improvement, the physical mechanism can be observed from the model outputs as well. In future work, the neural network is expected to be further applied in some other space weather studies (e.g., Dst, solar flare, etc).
Degree Doctor of Philosophy (PhD)
Institution RMIT University
School, Department or Centre Science
Subjects Space and Solar Physics
Simulation and Modelling
Photogrammetry and Remote Sensing
Keyword(s) Machine learning
Ionosphere
GNSS radio occultation
Modelling
Neural network
Versions
Version Filter Type
Access Statistics: 26 Abstract Views, 5 File Downloads  -  Detailed Statistics
Created: Tue, 24 Mar 2020, 08:28:01 EST by Keely Chapman
© 2014 RMIT Research Repository • Powered by Fez SoftwareContact us