17:15 〜 19:15
[PEM14-P12] 機械学習を利用したスプレッドFの検出とその応用
The Earth's ionosphere extends from an altitude of 50-60 km to 1000 km. In this region, atmospheric molecules and atoms of nitrogen and oxygen contained in the Earth's atmosphere are partially ionized by radiation energy such as ultraviolet radiation from the sun, and exist as plasma. One of the major characteristics of the ionosphere is its ability to reflect radio waves, and the maximum frequency at which radio waves are reflected is called foF2. This value is determined by the electron density.
Currently, as a preliminary study, there is a system that automatically reads foF2 when no ionospheric disturbances occur, based on the ionosonde observation installed at Shigaraki MU Observatory. This system functions effectively when the ionosphere is stable but is difficult to apply when ionospheric disturbances such as spread F occur. Therefore, to analyze ionograms during spread F events, Mask R-CNN, a machine learning model that excels in detecting objects in images, is utilized to learn ionograms during spread F occurrences. This enables ionograms affected by disturbances to be excluded from foF2 reading, allowing for highly accurate analysis.
When the Mask R-CNN model was trained on ionograms observed by the Shigaraki ionosonde during spread F occurrences, a detection accuracy of over 90% was achieved. Based on this, the same method was applied to detect spread F in ionograms observed in Chiang Mai, Thailand, for comparison and analysis with methods already in use at NICT, as well as for application to forecasting the occurrence of spread F. The results showed that a certain level of detection accuracy was also achieved for the Chiang Mai ionograms.
In addition, to evaluate the applicability of the model to different seasons and solar activity, validation using long-term ionogram data was conducted. Future work includes training with data from more diverse regions to improve the generalization performance of the model. In addition to Mask R-CNN, we will also utilize time series models such as LSTM and Transformer to improve the accuracy of spread F detection and to achieve real-time analysis.
Currently, as a preliminary study, there is a system that automatically reads foF2 when no ionospheric disturbances occur, based on the ionosonde observation installed at Shigaraki MU Observatory. This system functions effectively when the ionosphere is stable but is difficult to apply when ionospheric disturbances such as spread F occur. Therefore, to analyze ionograms during spread F events, Mask R-CNN, a machine learning model that excels in detecting objects in images, is utilized to learn ionograms during spread F occurrences. This enables ionograms affected by disturbances to be excluded from foF2 reading, allowing for highly accurate analysis.
When the Mask R-CNN model was trained on ionograms observed by the Shigaraki ionosonde during spread F occurrences, a detection accuracy of over 90% was achieved. Based on this, the same method was applied to detect spread F in ionograms observed in Chiang Mai, Thailand, for comparison and analysis with methods already in use at NICT, as well as for application to forecasting the occurrence of spread F. The results showed that a certain level of detection accuracy was also achieved for the Chiang Mai ionograms.
In addition, to evaluate the applicability of the model to different seasons and solar activity, validation using long-term ionogram data was conducted. Future work includes training with data from more diverse regions to improve the generalization performance of the model. In addition to Mask R-CNN, we will also utilize time series models such as LSTM and Transformer to improve the accuracy of spread F detection and to achieve real-time analysis.