Accurate and robust full-disk solar flare detection is crucial for space-weather forecasting and solar physics research. Traditional methods relying on single-band intensity and morphology suffer from high false-positive rates and fail to exploit the rich physical diagnostics within full-disk profiles. To address these limitations, this work proposes a full-disk flare detection framework that combines physics-driven dimensionality reduction with frequency-guided attention, tailored to multi-channel CHASE H $\alpha $ observations. High-quality full-disk flare masks and corresponding detection labels are constructed through disk cropping, limb darkening correction, clustering-based segmentation, and expert refinement. For spectral feature construction, each 118-channel H $\alpha $ line profile is modeled with a single Gaussian to extract line-core intensity, Doppler velocity, and FWHM, forming a physically interpretable multi-channel “fitted-image” input to the detector. Comparative experiments show that grayscale, PCA-based, autoencoder-reconstructed, and manually selected multi-channel images all perform noticeably worse than the fitted-image representation, whereas the latter yields clear gains in both AP50 and AP ${50\text{--}95}$ . Building on this representation, incorporating a custom frequency-guided module AdaFreq, MANet from Hyper-YOLO, and Focal-EIoU into YOLO11 further improves the best model’s AP50 and AP ${50\text{--}95}$ by about 3% and 7% respectively, over the baseline, and achieves superior performance compared withover the state-of-the-art detectors. These results demonstrate an effective, physics-constrained pathway for automated full-disk solar flare detection.
