|
ИСТИНА |
Войти в систему Регистрация |
ИСТИНА ПсковГУ |
||
DEVELOPMENT OF DENSE-FULLY CONVOLUTIONAL NETWORK FOR IONOGRAMSстатья Краткое сообщение
- Авторы: Enkhtuya Tsogtbaatar, Yung-Hui Li, A V.Dmitriev
- Сборник: Proceedings of the 18th Annual Meeting of the Asia Oceania Geosciences Society
- Год издания: 2021
- DOI: 10.1142/9789811260100_0067
- Аннотация: We apply algorithms of segmentation based on deep learning for monitoring the upper atmosphere and space weather by ionozonde product of ionograms. This machine learning technique has impressive independentadaptability consequences for fast and robust erasing of the noise from ionospheric signals that makes possible complete recovery of ionograms. The technique provides recognition systems and straight influence the ionograms confirmation, feature map visualization and classification results. This is the most significant and very challenging issue in this field. We used segmentation algorithm based on popular technology of convolutional neural networks, named as dense-fully convolutional network (DFCN) algorithm. We apply Taiwan data set of 6131 ionograms for training. Also, we accurately selected 1226 out of 6131 ionograms for testing of models. The DFCN model achieves the accuracy of 91% for E layer (IoU = 83.6%), 89% for F2 layer (IoU = 81%). These high-level accuracies outperform the existing approaches on challenging the ionogram data bases. The work was performed as part of NCU AI group: Li Y.-H., Tsogtbaatar E., Dmitriev A., Chang Y.-C., Hsieh M.-C., Hsu H.-W., Huang G.-H., Lin C.-H., Lin Y.-C., Mendoza M., Tsai L.-C.
- Добавил в систему: Дмитриев Алексей Владимирович