Shape-Preserving Embedding Technique for Binary Classification of Video Image of the Solar Surface

We study the embedding technique on the binary classification of video images as the explanatory variable.In this study, we assume the shape on video frame image has high sparsity and strong characteristic time evolution.In the embedding process, 2-dimensional image is resized keeping shape characteristics of the image and converted to a vector.The embedding allows dimensionality reduction from a 3-dimensional array (video image) as input data for machine learning to a 2-dimensional array of time sequences of embedded vectors.Using solar surface video images in the space weather field, we present evaluation experiments on multiple models with different embedding sizes, transformation formulas, and number of layers in the CNN.