Coronal mass ejection (CME) is a highly energetic solar phenomenon. It has a significant impact on the space weather in the near-Earth environment. With the accumulation of CME observations, it becomes more challenging to handle them manually. Therefore, we need an automatic method for identifying CMEs. We propose an unsupervised method for classifying and detecting changes in CMEs. The method consists of four main steps: (i) feature extraction: features derived from difference-image and features derived from pretrained convolutional neural networks (CNN), (ii) dimensional reduction using Principal Component Analysis (PCA), (iii) unsupervised classification using K-mean clustering based on PCA components and (iv) morphological post-processing to improve the clustering output. We compare the results with manual catalog (e.g., coordinated data analysis workshops (CDWA) data center) and automatic detection catalogs (e.g., solar eruption detection system (SEEDS), computer-aided CME tracking (CACTus) and coronal image processing (CORIMP)). The comparison is based on CME characteristics (e.g., time of first appearance, position angle, angular width and velocity). We demonstrate the benefit of this unsupervised method, which produces comparable results to classical methods.

日冕物质抛射（CME）是一种高能量的太阳现象。它对近地环境中的空间天气有重要影响。随着 CME 观察的积累，手动处理它们变得更具挑战性。因此，我们需要一种自动识别 CME 的方法。我们提出了一种用于分类和检测 CME 变化的无监督方法。该方法由四个主要步骤组成：(i) 特征提取：从差异图像派生的特征和从预训练卷积神经网络 (CNN) 派生的特征，(ii) 使用主成分分析 (PCA) 进行降维，(iii) 无监督分类使用基于 PCA 组件的 K 均值聚类和 (iv) 形态学后处理来提高聚类输出。我们将结果与手动目录（例如，协调数据分析研讨会 (CDWA) 数据中心）和自动检测目录（例如，太阳爆发检测系统 (SEEDS)、计算机辅助 CME 跟踪 (CACTus) 和日冕图像处理 (CORIMP)）。该比较基于 CME 特性（例如，首次出现的时间、位置角、角宽度和速度）。我们证明了这种无监督方法的好处，它产生与经典方法相当的结果。