MMFNet-K: A multimodal data fusion network for nondestructive freshness assessment of multiple kiwifruit cultivars

Kiwifruit freshness changes rapidly post-harvest, yet subtle external visual manifestations of physiological deterioration and complex environments hinder precise unimodal assessment. This study proposes MMFNet-K, a Multimodal Fusion Network for kiwifruit freshness classification integrating RGB imaging, hyperspectral imaging (HSI), and electrochemical impedance spectroscopy (EIS). The model utilizes a dual-branch architecture with two-stage fusion involving feature-level and decision-level strategies to achieve deep coupling of cross-modal information. To validate generalization, two logistics scenarios including standard cold chain (4 °C) and ambient retail (25 °C) were simulated across three major cultivars including Hayward, Golden, and Xuxiang to construct a non-destructive dataset. Objective freshness grading standards were established and verified through firmness and soluble solids content analysis. Experimental results demonstrate that MMFNet-K achieved accuracies ranging from 95.6%-99.3% across all six cultivar-temperature combinations, with a macro-average precision of 0.978, a recall of 0.977, and an F1-score of 0.977. Comparative analysis indicates that the model outperforms unimodal methods by 9.13% and surpasses existing mainstream multimodal fusion schemes by 5.43%. These results demonstrate the high robustness of MMFNet-K in complex post-harvest environments and its significant competitive advantage over traditional methods. Furthermore, a practical application scheme was designed based on this model to provide a feasible technical path for kiwifruit freshness grading and contribute to the reduction of waste and losses during post-harvest processes.