A Hybrid Framework for Automated Sunspot Recognition and LeNet-5-Based Validation Using Multiple Solar Observations

Automated sunspot recognition is critical for understanding solar activity and its space weather impacts. In this study, a recognition-verification framework that integrates a automated sunspot recognition with a deep learning-based classifier were developed. Sunspots were initially recognized using a combination of threshold segmentation and mathematical morphology applied to SOHO/MDI full-disk white-light intensity continuum images spanning the entire solar cycle 23. To verify the accuracy of recognition algorithm, an improved LeNet-5 network was trained on a balanced dataset of sunspot and background samples, achieving classification accuracies of 99.67\% and 99.76\% on validation and testing sets, respectively. Subsequently, sunspot features, including number and area, were extracted and compared with the solar region summary records. The Spearman and Pearson correlation coefficients for sunspot number were 0.8710, 0.8281, respectively, and those for sunspot area reached 0.9386 and 0.9368, confirming the reliability of the recognition results. In addition, to evaluate generalizability of the proposed framework, it was applied to the reduced-resolution SDO/HMI images from the overlapping period with SOHO/MDI during solar cycle 24. Classification accuracies remained above 99\%, and feature consistency between instruments was high. However, correlation coefficients during the overlapped period of solar cycle 24 were notably lower than those from the full solar cycle 23, which may be attributed to differences in the solar cycle phase and observational conditions. These findings demonstrate the robustness and applicability of the proposed hybrid framework for sunspot recognition and verification.