A Deep Learning Approach for Automated Total Sunspot Number Estimation

Abstract Accurate sunspot number estimation is essential for understanding the long-term evolution of solar activity and its impact on space weather. Sunspot numbers have been manually determined, leading to inconsistencies and observer-dependent biases. To address this, the World Data Center Sunspot Index and Long-term Solar Observations (WDC-SILSO) aggregates data from a global network of observatories to estimate the daily total sunspot number, enabling cross-validation and calibration across simultaneous observations. This study proposes a novel deep learning framework for automated total sunspot number calculation using solar full-disk continuum images from the Solar Dynamics Observatory. The method integrates U-Net for sunspot segmentation, K-means clustering for distinguishing umbrae from penumbrae, and You Only Look Once model for sunspot group detection. The selection of image-processing thresholds and neural network hyperparameters is optimized with respect to WDC-SILSO reference values during training. The results demonstrate a high correlation of 0.97 between the estimated and the WDC-SILSO daily total sunspot numbers. Furthermore, the framework offers a scalable approach suitable for future high-resolution solar observations.