Beef freshness classification based on RGB–hyperspectral fusion and temperature-aware temporal modeling

Rapid and accurate assessment of beef freshness is essential for ensuring food safety and quality control. However, existing methods predominantly rely on single-modal sensing and fail to adequately capture the temporal dynamics of the spoilage process. To address these limitations, this study proposes a multimodal framework that integrates RGB imaging and hyperspectral imaging with temperature-aware temporal modeling for beef freshness classification. A dual-branch tokenization module is developed to extract spatial features from RGB images and spectral discriminative features from hyperspectral data. Subsequently, a complementary perception-aware fusion (CPAF) module is introduced to achieve effective multimodal feature integration through feature alignment and discrepancy modeling, improving the F1-score from 0.8393 to 0.9461 compared with a simple concatenation baseline. Furthermore, a temperature-aware two-layer GRU is employed to model temporal dependencies across multiple time points, resulting in a further increase of the F1-score to 0.9961 and an accuracy of 99.64%, outperforming HybridSN ( F 1 = 0.9922 ) and 3D-CNN ( F 1 = 0.9715 ). Experimental results on beef samples collected at multiple time points under room temperature demonstrate that the integration of multimodal fusion and temporal modeling significantly enhances classification performance. The proposed method provides an effective approach for rapid and non-destructive beef freshness assessment.