11:00 AM - 1:00 PM
[PEM12-P17] Comparison of general object detection models for Ionogram sporadic E layer echoes detection
Keywords:Ionogram, sporadic E layer, general object detection, supervised machine learning
Since the appearance of sporadic E layers in the ionosphere causes interference on radio communications, it is important in space weather forecast to detect the occurrence of sporadic E layers automatically and continuously from ionogram images in quasi-real time.The purpose of this paper is to propose a highly accurate and robust method for detecting ionospheric disturbance sporadic E-layer echoes from ionogram images. We adopt some existing popular general object detection methods based on convolutional neural networks (CNN), for our approach to detect sporadic E layer echoes as object instances in an Ionogram image. The main objective of this paper is to clarify the accuracy of sporadic E layer detection and most useful framework model for three well-known object detection models, Faster-RCNN, YOLO and SSD. Each model consists of deep neural networks to detect objects by proposing regions and classifying them. Faster-RCNN has two separate networks, one network is for searching candidate regions and the other is for classifying them, while YOLO and SSD perform their search and classification in a unified network. Faster-RCNN is not able to detect in real time, YOLO is less likely to make mistakes between objects and background, and SSD is able to detect even small objects withhigh accuracy.
In this study, we prepare 1178 ionogram images (January-December 2019, Sasaguri in Japan) and split them randomly into the training data (942 images) and the validation data (236 images) in the ratio of 8:2 for our experiments. We apply two preprocessing process, noise reduction using the bilateral filter which is one of smoothing filters and smoothing in the horizontal direction, into each ionogram image. We also perform the annotation process for the training data and give the position information of bounding boxes as the object regions and their class label.
The results show that the Faster-RCNN and YOLO methods can detect sporadic E-layer echoes with high accuracy of 98.96% AP and98.68% AP, respectively, compared to 88.67% AP for SSD (AP is Average Precision: the most commonly used metric is AP for the accuracy of object detection network, derived from precision and recall). The average automatic scaling error of foEs was 0.1495, 0.2615, and 0.4373 MHz for Faster-RCNN, YOLO, and SSD, respectively.
This study reveals that the general object detection method of end-to-end learning is useful for echo region search and its classidentification as a two-class classification and regression problem in ionogram images. On the other hand, there is still room to reduce the computational complexity of the candidate region search by changing the architecture to match the characteristics of each echo, including not only the sporadic E-layer echoes but also the normal F-layer echoes or the spread-F echoes during ionospheric disturbances. In the future, we aim to extend the object detection model optimized by ionogram images as a multi-class classification andregression problem.
In this study, we prepare 1178 ionogram images (January-December 2019, Sasaguri in Japan) and split them randomly into the training data (942 images) and the validation data (236 images) in the ratio of 8:2 for our experiments. We apply two preprocessing process, noise reduction using the bilateral filter which is one of smoothing filters and smoothing in the horizontal direction, into each ionogram image. We also perform the annotation process for the training data and give the position information of bounding boxes as the object regions and their class label.
The results show that the Faster-RCNN and YOLO methods can detect sporadic E-layer echoes with high accuracy of 98.96% AP and98.68% AP, respectively, compared to 88.67% AP for SSD (AP is Average Precision: the most commonly used metric is AP for the accuracy of object detection network, derived from precision and recall). The average automatic scaling error of foEs was 0.1495, 0.2615, and 0.4373 MHz for Faster-RCNN, YOLO, and SSD, respectively.
This study reveals that the general object detection method of end-to-end learning is useful for echo region search and its classidentification as a two-class classification and regression problem in ionogram images. On the other hand, there is still room to reduce the computational complexity of the candidate region search by changing the architecture to match the characteristics of each echo, including not only the sporadic E-layer echoes but also the normal F-layer echoes or the spread-F echoes during ionospheric disturbances. In the future, we aim to extend the object detection model optimized by ionogram images as a multi-class classification andregression problem.