RU-net: A Residual U-net for Automatic Interplanetary Coronal Mass Ejection Detection

Abstract

Detection methods for interplanetary coronal mass ejections (ICMEs) from in situ spacecraft measurements are mostly manual, which are labor-intensive and time-consuming, being prone to the inconsistencies of identification criteria and the incompleteness of the existing catalogs. Therefore, the automatic detection of ICMEs has aroused the interest of the astrophysical community. Of these automatic methods, the convolutional neural network–based methods show the advantages of fast speed and high precision. To further improve the computing feasibility and detection performance, this paper proposes a method called residual U-net (RU-net), from the perspective of time-series segmentation. With the help of U-net architecture, we design an encoder–decoder network with skip connection to capture multiscale information, where the end-to-end architecture with an embedded residual element is formulated to accelerate the algorithmic convergence. For the in situ data from 1997 October 1 to 2016 January 1 collected by the Wind spacecraft, the results of our experiments demonstrate the competitive performance of the proposed RU-net in terms of accuracy and efficiency (178 of 230 ICMEs are detected in the test set, and the F1 score is 80.18%).

Automatic Detection of Interplanetary Coronal Mass Ejections in Solar Wind In Situ Data

“…The average Dice Coefficient (also known as F1 score) achieved by Chen et al. (2022) is 0.68, which is comparable to our findings. However, to thoroughly compare both pipelines, a reimplementation and subsequent cross validation of the pipeline proposed by Chen et al.…”

“…Our pipeline is comparable to the one proposed by Chen et al. (2022). Nevertheless, we use a different validation method focusing on generalization, as conventional within the machine learning community to avoid overfitting and predict the performance on new unseen data more accurately.…”

“…Nevertheless, its combination with other forecasting tools can already be expected to serve as a significant contribution in the area of space weather forecast. Our pipeline is comparable to the one proposed by Chen et al (2022). Nevertheless, we use a different validation method focusing on generalization, as conventional within the machine learning community to avoid overfitting and predict the performance on new unseen data more accurately.…”

“…In this article, we rephrase the automatic detection of ICMEs as a time series segmentation problem and propose a pipeline using a variation of the ResUNet++ (Jha et al, 2019), which shows improvements compared to an existing method. A comparable study was published during the finalization of this manuscript by Chen et al (2022). Nevertheless, we use a different validation method focusing on generalization, as conventional within the machine learning community to avoid overfitting and predict the performance on new unseen data more accurately.…”

DeepBSA: A deep-learning algorithm improves bulked segregant analysis for dissecting complex traits

“…Nine models were trained with different number of bulked pools from two to 10. In the encoder, the features are extracted by four repeated applications of residual operation (Chen et al, 2022) and max pooling operation with stride 2 for downsampling. After each downsampling, the number of feature channels are doubled and denoted on top of the box.…”