VLF Remote Sensing of the D Region Ionosphere Using Neural Networks

Abstract

Narrowband very low frequency (VLF) remote sensing has proven to be a useful tool for characterizing the ionosphere's D region (60-to 90-km altitude) electron density. This work expands upon single transmitter-receiver pair electron density profile inference methods to create a more generalized narrowband VLF remote sensing method that concurrently resolves a two-parameter electron density profile for an arbitrary number of transmitter-receiver pairs. A target function is constructed to take in a single time step of narrowband amplitude and phase observations from an arbitrary number of transmitter and receiver combinations and return the inferred average waveguide parameters along all paths. The target function is approximated using an artificial neural network (ANN). Synthetic training data are generated using the U.S. Navy's Long-Wavelength Propagation Capability program, which is then used to train the ANN. Performance of the ANN with real-world data is measured in two ways. First, ANN inferred average waveguide parameters are compared to a variety of previously published narrowband VLF remote sensing experiments. Second, ANN inferred average waveguide parameters are used in Long-Wavelength Propagation Capability to predict narrowband VLF amplitude and carrier phase at a receiver that was withheld when performing the average waveguide parameter inference. Results show the approximated target function performs well in capturing temporal and spatial characteristics of the D region.

Data‐Driven Forecasting of Low‐Latitude Ionospheric Total Electron Content Using the Random Forest and LSTM Machine Learning Methods

“…Forecasting of geomagnetic indices such as the Kp and Dst (Tan et al, 2018;Wu & Lundstedt, 1996), coronal mass ejection propagation time (Bobra & Ilonidis, 2016), solar wind speed (Yang et al, 2018), relativistic electrons at geosynchronous orbits (Ling et al, 2010) have been achieved by applying machine learning methods. Other recent works by Bortnik et al (2018), Z. Chen, et al (2019, McGranaghan et al (2018), and Gross and Cohen (2020) applied different machine learning methods for different space weather studies over the ionosphere, magnetosphere and the radiation belt. A non-linear regression analysis was used by Villalobos and Valladares (2020) to model TEC values over South and Central America.…”

“…Other recent works by Bortnik et al (2018), Z. Chen et al (2019), McGranaghan et al (2018), and Gross and Cohen (2020) applied different machine learning methods for different space weather studies over the ionosphere, magnetosphere and the radiation belt. A non-linear regression analysis was used by Villalobos and Valladares (2020) to model TEC values over South and Central America.…”

Seasonal Variation of the D‐Region Ionosphere: Very Low Frequency (VLF) and Machine Learning Models

“…Since the properties of the synthetic ionospheres are known, we can compute the average h' and  E for each path to be compared with the model's output during training. were set based on previous studies of the daytime ionosphere (Gross & Cohen, 2020;Thomson, 1993;Thomson & McRae, 2009). More details of the synthetic ionosphere generation process are available in Gross and Cohen (2020).…”

“…The overall model training process is described in Figure 2. This training process is also used by Gross and Cohen (2020).…”

“…A 3 × 3 median filter is applied to the 2D matrix of seed errors to remove spurious outliers, and the seed with the lowest error is then selected as the optimal seed. For the majority of days, the median filter is not necessary and was not used in the original design by Gross and Cohen (2020); however, occasionally, an isolated global minimum seed error occurs as shown by the red circles in Figure 3. We expect nearby seeds to have similar performance since a small change in seed value corresponds to a similarly small change in the reference value.…”