5:15 PM - 6:30 PM
[PEM09-P23] Detection of Pi2 geomagnetic pulsation by machine learning and development of real-time substorm warning system
Keywords:machine learning, convolutional neural network, Pi2 pulsations, aurora, substorm
Pi2 geomagnetic pulsations are defined to have an irregular waveform with a period of 40-150 seconds. It is known that they appear during auroral breakups and intensifications, and the occurrence of Pi2 geomagnetic pulsations can be used to estimate the occurrence of auroral breakups, even at low latitudes where auroras cannot be seen directly. If Pi2 geomagnetic pulsations are detected in real time, we can nowcast disturbances in the polar regions where auroras occur and disturbances in the magnetosphere that provides the excitation energy of auroras. In this study, machine learning is used to automatically detect Pi2 pulsations from low latitude geomagnetic data.
The 1-sec geomagnetic field data are acquired at the Inabu observatory in Aichi prefecture. Geomagnetic coordinates of the Inabu observatory are 26.77°N and 207.49°E. We prepare 3000-5500 images by plotting the geomagnetic field data for the month of February 2020 and use them as training dataset for the machine learning. Visually scanning these images, we label them occurrence/no occurrence of Pi2 pulsations. Using these labeled images, we train a convolutional neural network (CNN), one of types of machine learning, as it can determine whether Pi2 pulsations are present or not in images. As CNN models, we use our own model and Resnet50. It is revealed that when the training dataset includes the equal numbers of occurrence/no occurrence of Pi2 images, CNN obtains a better learning result. The best performance is achieved by ResNet50 with 93% accuracy.
The future plan is to transfer geomagnetic field data continuously in real time from Inabu observatory to Nagoya University and to apply the developed CNN model to detect Pi2 pulsations in an automated way. By doing so we will construct a realtime substorm-warning system.
The 1-sec geomagnetic field data are acquired at the Inabu observatory in Aichi prefecture. Geomagnetic coordinates of the Inabu observatory are 26.77°N and 207.49°E. We prepare 3000-5500 images by plotting the geomagnetic field data for the month of February 2020 and use them as training dataset for the machine learning. Visually scanning these images, we label them occurrence/no occurrence of Pi2 pulsations. Using these labeled images, we train a convolutional neural network (CNN), one of types of machine learning, as it can determine whether Pi2 pulsations are present or not in images. As CNN models, we use our own model and Resnet50. It is revealed that when the training dataset includes the equal numbers of occurrence/no occurrence of Pi2 images, CNN obtains a better learning result. The best performance is achieved by ResNet50 with 93% accuracy.
The future plan is to transfer geomagnetic field data continuously in real time from Inabu observatory to Nagoya University and to apply the developed CNN model to detect Pi2 pulsations in an automated way. By doing so we will construct a realtime substorm-warning system.