3:30 PM - 5:00 PM
[PEM15-P04] Automatic detection and statistical analysis of sporadic E layer by machine learning using Shigaraki ionosonde observation
Keywords:Sporadic E, Machine Learning, Ionosphere
In Shigaraki MU Observatory, where the MU radar is a main facility, an ionosonde has been operated since the beginning of the MU radar operation in 1984. Digital images after 2001 are available in the database of Research Institute for Sustainable Humanosphere (RISH), Kyoto University ( http://database.rish.kyoto-u.ac.jp/arch/mudb/ionosonde/ ). However, the scaled parameters of the Shigaraki ionograms have not been available in the database, so that users have to scale the ionograms by their own. In order to utilize the Shigaraki ionosonde easily, we have been developing an auto-scaling system of ionograms. A representative model of instance segmentation, which is called Mask Region-Convolutional Neural Network (R-CNN), is used as a background model. By using the machine learning model, foF2 can be successfully scaled automatically. In this research, we aim to detect sporadic-E (Es) layer from ionograms and auto-scale foEs values.
The system was trained by 512 ionogram images in which ionospheric traces of Es layer are labeled manually. The obtained foEs values were validated with other 150 images. We confirmed a high accuracy of the auto-scaled foEs values with an averaged error of as low as 0.24MHz. Comparing with Kokubunji ionosonde operated by National Institute of Information and Communications Technology (NICT), whose distance from Shigaraki ionosonde is about 350 km, a distribution of foEs values are similar with that from Kokubunji ionosonde, which implies the auto-scaling works correctly. However,one-to-one comparison of foEs obtained in the two stations at the same time show a relatively low correlation coefficient (0.51). It suggests that a large-scale formation of Es layer is common over Shigaraki and Kokubunji, while the distribution of a neutral wind and metallic ions have a different pattern over the two stations. From a long-term analysis, we did not find a dependence on solar activity.
The system was trained by 512 ionogram images in which ionospheric traces of Es layer are labeled manually. The obtained foEs values were validated with other 150 images. We confirmed a high accuracy of the auto-scaled foEs values with an averaged error of as low as 0.24MHz. Comparing with Kokubunji ionosonde operated by National Institute of Information and Communications Technology (NICT), whose distance from Shigaraki ionosonde is about 350 km, a distribution of foEs values are similar with that from Kokubunji ionosonde, which implies the auto-scaling works correctly. However,one-to-one comparison of foEs obtained in the two stations at the same time show a relatively low correlation coefficient (0.51). It suggests that a large-scale formation of Es layer is common over Shigaraki and Kokubunji, while the distribution of a neutral wind and metallic ions have a different pattern over the two stations. From a long-term analysis, we did not find a dependence on solar activity.