In the non-destructive detection of soluble solids content (SSC) across multi-category fruits, hyperspectral imaging (HSI) often faces challenges such as dynamic localization distortion of biological feature regions and significant inter-species optical heterogeneity. This study proposes HyperVision-HSI, a classification-guided spectral-spatial adaptive decoupling framework. By integrating dynamic ROI localization, dual-channel spectral calibration, and a category-aware model invocation architecture, the framework achieves precise feature extraction and matching for multi-category samples. Utilizing real-time classification information as a decision-making shunt, the framework automatically triggers species-adaptive thresholds to extract high-purity ROIs for grapes, tomatoes, and Xiangli pears—three species with markedly different optical properties. It then dynamically invokes pre-optimized specialized regression models (Grapes: KRR; Tomatoes: BRR; Xiangli Pears: Lasso), effectively addressing the feature dilution problem encountered by single models when processing heterogeneous samples. Experimental results demonstrate that the system achieves SSC prediction errors (RMSE) as low as 0.62°Brix, 0.32°Brix, and 0.37°Brix on independent test sets for the three fruit types, respectively, with a single-frame processing time of less than 1 second. The modular architecture and high scalability of HyperVision-HSI provide a rigorous adaptive technical pathway for the automated detection of multi-category agricultural phenomics in diverse scenarios.
Dong et al. (Wed,) studied this question.