Vision-Based CNN Prediction of Sunspot Numbers from SDO/HMI Images

Sunspot numbers constitute the longest and most widely used record of solar activity, with direct implications for heliophysical research and space-weather applications. Traditional sunspot counting relies on visual inspection and algorithmic feature-detection pipelines, which can be affected by observer-dependent choices, image quality, and methodological variability across implementations. Recent advances in deep learning, particularly convolutional neural networks (CNNs), enable the direct use of solar imagery for automated image-to-scalar regression, reducing the need for explicit, handcrafted feature design. In this work, we present a supervised vision-based framework to estimate the daily sunspot number from full-disk continuum images acquired by the Helioseismic and Magnetic Imager (HMI) onboard NASA’s Solar Dynamics Observatory (SDO). Images from 2011–2024 were paired with daily sunspot numbers from the SILSO Version 2.0 dataset maintained by the Royal Observatory of Belgium. After preprocessing and augmentation, a CNN was trained to infer a scalar sunspot number directly from pixel data at the observation time of each image. The proposed model achieved strong performance on an independent test split (\(R^2\)=0.964, RMSE=9.75, MAE=6.74), indicating close agreement with SILSO reference values across a broad activity range. In comparison with prior studies, we position this approach as a competitive and conceptually simple alternative for direct image-based estimation, complementing time-series forecasting models that target monthly means or smoothed indices. Interpretability analyses using Grad-CAM and Integrated Gradients indicate that the network consistently attributes relevance to sunspot-bearing regions when forming its estimates. These results highlight the potential of deep vision-based approaches for scalable solar monitoring and automated estimation of classical heliophysical indices. Future work should explore multimodal fusion with additional observables (e.g., magnetograms) and standardized cross-cycle benchmarks to strengthen robustness under changing solar conditions.