Title and Abstract

Title:  ML-Eco Fi: A Machine Learning Ecosystem for Automatic Analysis of Solar Filaments

Abstract:  Building an effective system for automatic analysis of scientific objects (e.g., physical phenomenon) is far more complex than running a Deep Neural Network algorithm utilizing a powerful cluster of GPUs. It imposes numerous challenges on multiple fronts, many of which rarely find their ways into general-purpose computer vision research. MLEco Fi is an ongoing, NSF-sponsored project that aims at building an ecosystem of data, tools, and techniques, which brings the indisputable power of machine learning algorithms at the fingertips of domain experts. Moreover, it takes one big step towards properly tackling scientific object-detection problems such as those needed for microscopic, satellite, aerial, X-ray, or telescopic imagery. To ensure the effectiveness of our products, we teamed up with the National Solar Observatory to provide us with their wealth of knowledge and expertise. This ecosystem targets solar filaments. Filaments are clouds of ionized gas (plasma) in the solar chromosphere, which are critical for Space Weather Forecasting, as they can tell us about the occurrence of geomagnetic storms. The most important piece of information we want to extract from the detected filaments is the directions of their axial field on the Sun. In this talk, I will discuss our team's plan as we progress in the development of ML-Eco Fit, as well as the challenges on the way. In the process, I will review the methods and concepts I have introduced in the past few years towards this goal. These efforts involve the manual annotation of solar filaments captured by the GONG's ground-based observatories, evaluation of the manual annotation pipeline, devising object-similarity measures sensitive to the fine structures of filaments, augmentation of filaments, and evaluation of models' learning process from imbalanced data.