Vision-based CNN prediction of sunspot numbers from SDO/HMI images

Abstract

Sunspot numbers provide the longest continuous record of solar activity and remain a key index for heliophysical research and space-weather applications. Standard sunspot determination relies on visual inspection and algorithmic feature-detection pipelines, both of which involve methodological choices and can be sensitive to image quality and implementation details. Convolutional neural networks (CNNs) offer an alternative by learning an end-to-end mapping from solar images to a scalar index, reducing reliance on explicit, handcrafted feature design. Here we present a supervised vision-based regression framework to estimate the daily sunspot number from full-disk continuum images acquired by the Helioseismic and Magnetic Imager (HMI) onboard NASA Solar Dynamics Observatory (SDO). We pair daily images from 2011-2024 with the SILSO Version 2.0 daily sunspot number and train a CNN to infer the scalar value at the observation time of each image. On an independent test split, the model achieves R2=0.964, RMSE=9.75, and MAE=6.74, indicating close agreement with SILSO across a wide activity range. Interpretability analyses using Grad-CAM and Integrated Gradients show that the network attributions concentrate on sunspot-bearing regions, supporting the physical plausibility of the learned representations. These results demonstrate the feasibility of direct image-to-index estimation for scalable solar monitoring. Future work will explore multimodal fusion with complementary observables (e.g., magnetograms) and standardized cross-cycle benchmarks to assess robustness under changing solar conditions.