An High Resolution and Efficient System for Labeling Solar Coronal Hole Images

Abstract

The ClassX project is based on the development of an automatic training dataset labeling tool and online service to fill the gap of missing high-quality training image datasets. Novel spatiotemporal AI/ML-based capabilities are being developed to automatically classify, label, store, and share training datasets among a group of needed users. It is currently being expanded to the heliophysics domain. In the field of solar physics, accurately identifying and labeling coronal holes in solar images is crucial for understanding solar dynamics and space weather prediction. However, manual labeling of these features is time-consuming and tedious. To address this, we present an automated image labeling tool that leverages the Mask Region-based Convolutional Neural Network (Mask R-CNN) model to detect and generate masks for coronal holes in solar images. We trained this model on a dataset of solar images curated by segmenting images from NASA’s Solar Dynamics Observatory using ClassX’s Quickshift Segmentation implementation and then labeling them manually. Our results demonstrated that the Mask R-CNN model achieved acceptable performance in coronal hole labeling. Overall, accuracy of this model compares favorably to the U-NET model. This performance underscores the potential of implementing Mask R-CNN in reducing the manual labor required for dataset labeling and ensuring consistency across large datasets. The successful implementation of these models in labeling coronal holes will not only advance the field of solar physics but also highlight the broader applicability of ClassX in other scientific domains, fostering the development of precise and reliable machine-learning models for scientific and engineering purposes.