A Machine Learning Ecosystem for Filament Detection: Classification, Localization, and Segmentation

<p dir="ltr">Over the past decade, object-detection algorithms have achieved remarkable success, but they remain limited when applied to scientific imagery such as telescopic, satellite, and medical images. Challenges like pixel-precise segmentation and capturing fine-grained features persist, particularly in interdisciplinary domains like Space Weather. This project focuses on solar filaments, which are key to forecasting extreme solar events that can disrupt power grids and GPS systems. We present a Machine Learning Ecosystem that enables automated, accurate analysis of filament dynamics. It includes two data products: the largest manually annotated filament dataset and a growing set of automatically labeled images, and four software tools for filament localization, segmentation, and classification. In collaboration with the National Solar Observatory, we utilize data from the GONG network, which offers continuous solar imaging from six ground-based observatories. In addition to advancing solar forecasting, this project contributes datasets and tools useful to the broader computer vision community, particularly for high-precision segmentation in scientific contexts.<br></p>