Steps Toward a Large-Scale Solar Image Data Analysis to Differentiate Solar Phenomena

We detail the investigation of the first application of several dissimilarity measures for large-scale solar image data analysis. Using a solar-domain-specific benchmark dataset that contains multiple types of phenomena, we analyzed combinations of image parameters with different dissimilarity measures to determine the combinations that will allow us to differentiate between the multiple solar phenomena from both intra-class and inter-class perspectives, where by class we refer to the same types of solar phenomena. We also investigate the problem of reducing data dimensionality by applying multi-dimensional scaling to the dissimilarity matrices that we produced using the previously mentioned combinations. As an early investigation into dimensionality reduction, we investigate by applying multidimensional scaling (MDS) how many MDS components are needed to maintain a good representation of our data (in a new artificial data space) and how many can be discarded to enhance our querying performance. Finally, we present a comparative analysis of several classifiers to determine the quality of the dimensionality reduction achieved with this combination of image parameters, similarity measures, and MDS.