5:15 PM - 7:15 PM
[PEM14-P13] Long-term Statistical Analysis of Shigaraki Ionosonde Observations Using Machine Learning Models
Keywords:Ionosphere, Ionosonde, MU Radar, Machine Learning
The Earth's ionosphere exists at altitudes ranging from approximately 50–60 km to 1000 km. In this region, molecules and atoms are partially ionized by radiation energy such as ultraviolet rays from the Sun, generating plasma. The ionosphere has the property of reflecting radio waves, and the reflection frequency and reflection altitude are determined by the altitude distribution of electron density. Therefore, long-term observation and analysis of electron density in the ionosphere play an important role in the development of communication technology.
There are several methods for ionospheric observation, including ionosondes, incoherent scatter (IS) radars, and satellites, each with its own characteristics. The MU Observatory, located in Shigaraki, Koka City, Shiga Prefecture, is equipped with an ionosonde and the MU radar, an atmospheric observation radar for studying the middle and upper atmosphere. The MU radar has been used as an IS radar for ionospheric observation. However, the absolute value of electron density cannot be directly obtained from MU radar data, requiring calibration using ionospheric parameters obtained from ionograms. Currently, while the ionosonde at the Shigaraki MU Observatory continues to observe the ionosphere, the extraction of ionospheric parameters from its data is not performed manually. Furthermore, manually reading ionospheric parameters from all past observational data would require an enormous amount of time and effort, making it impractical.
To address this issue, this study implemented an automatic extraction system for the critical frequency of the F2 layer (foF2) using a machine learning model for ionosonde data interpretation. The main steps of the processing algorithm include noise removal and segmentation-based prediction using machine learning. In the noise removal process, vertical and horizontal noise present in the Shigaraki ionograms was reduced using fourth-order polynomial fitting and local averaging of adjacent values, respectively. For segmentation-based prediction using machine learning, labels were assigned to Fo and Fx, as well as to cases where fo and fx could not be distinguished due to spread F, denoted as F. Training was conducted using Mask R-CNN, a widely used architecture for instance segmentation.
The extracted foF2 values were compared with ionosonde observation data from Kokubunji to verify measurement validity. Additionally, comparisons were made with the IRI model, an empirical ionospheric model, to evaluate the long-term statistical characteristics at Shigaraki. The machine learning model was applied to data from 2002 to 2023, and a strong correlation was observed when comparing the values with those from Kokubunji. Moreover, the mean squared error exhibited dependencies on time and month, with a tendency for larger errors between April and August.
There are several methods for ionospheric observation, including ionosondes, incoherent scatter (IS) radars, and satellites, each with its own characteristics. The MU Observatory, located in Shigaraki, Koka City, Shiga Prefecture, is equipped with an ionosonde and the MU radar, an atmospheric observation radar for studying the middle and upper atmosphere. The MU radar has been used as an IS radar for ionospheric observation. However, the absolute value of electron density cannot be directly obtained from MU radar data, requiring calibration using ionospheric parameters obtained from ionograms. Currently, while the ionosonde at the Shigaraki MU Observatory continues to observe the ionosphere, the extraction of ionospheric parameters from its data is not performed manually. Furthermore, manually reading ionospheric parameters from all past observational data would require an enormous amount of time and effort, making it impractical.
To address this issue, this study implemented an automatic extraction system for the critical frequency of the F2 layer (foF2) using a machine learning model for ionosonde data interpretation. The main steps of the processing algorithm include noise removal and segmentation-based prediction using machine learning. In the noise removal process, vertical and horizontal noise present in the Shigaraki ionograms was reduced using fourth-order polynomial fitting and local averaging of adjacent values, respectively. For segmentation-based prediction using machine learning, labels were assigned to Fo and Fx, as well as to cases where fo and fx could not be distinguished due to spread F, denoted as F. Training was conducted using Mask R-CNN, a widely used architecture for instance segmentation.
The extracted foF2 values were compared with ionosonde observation data from Kokubunji to verify measurement validity. Additionally, comparisons were made with the IRI model, an empirical ionospheric model, to evaluate the long-term statistical characteristics at Shigaraki. The machine learning model was applied to data from 2002 to 2023, and a strong correlation was observed when comparing the values with those from Kokubunji. Moreover, the mean squared error exhibited dependencies on time and month, with a tendency for larger errors between April and August.