Speaker
Description
Ionospheric scintillation refers to rapid phase and amplitude fluctuations of radio signals as they pass through ionospheric irregularities. While commercial scintillation monitors have been used extensively to study scintillation, their relatively high costs have limited scientific use. To address this, we have been developing low-cost scintillation monitors based on single-board computers and Global Navigation Satellite System (GNSS) receivers. These instruments, referred to as ScintPi, have been deployed at different locations over the globe, producing a very large and valuable dataset which has motivated the development of data-driven processing and analysis techniques.
This work presents a two-stage machine learning (ML) framework applied to scintillation measurements made by ScintPi monitors. More specifically, the framework consists of: (1) a supervised classification to automatically identify scintillation in GNSS data, and (2) a supervised regression to infer scintillation severity from Total Electron Content (TEC) measurements.
In the first stage, the goal is to distinguish true ionospheric scintillation from non-geophysical disturbances such as multipath and radio frequency interference (RFI). Supervised classification models were trained using a labeled dataset with known disturbance sources: multipath, RFI, or low-latitude scintillation. Features were derived from signal power spectra and TEC. Preliminary results show that tree-based models can distinguish scintillation from multipath and RFI with approximately 99% classification accuracy. These results were used to generate a cleaned dataset of approximately 1,100,000 low-latitude, quiet-time scintillation observations during 2024.
In the second stage, the goal is to use the cleaned dataset from the first stage to interpret the Rate of Change of TEC Index (ROTI) derived from lower-rate GNSS observations. While ROTI is expected to be correlated with the amplitude scintillation index S4, the relationship between these two variables depends on many factors including link geometry, irregularity drift, and irregularity spectrum. Determining a relationship between S4 and ROTI is important since it could allow estimates of scintillation severity from the wide network of geodetic receivers rather than specialized scintillation monitors alone. Supervised regression models were trained to infer S4 from ROTI. Features include ROTI from 1 Hz TEC, link geometry, local time, and solar activity. Model performance was evaluated using an independent low-latitude station outside the longitude sector used for training. Preliminary results show strong agreement between estimated and observed S4 (Pearson r ~ 0.92).
Overall, the results demonstrate the usefulness of ML techniques to aid processing of scintillation measurements and to relate scintillation parameters to different types of observations (e.g., ROTI).