Multimodal Data Fusion Based Intelligent Grading System for Kiwifruit Maturity Using Spectral‐Impedance Properties

ABSTRACT The quality of kiwifruit directly affects its market value. However, current grading methods mainly rely on manual operations, with incomplete grading standards and limited accuracy. To address this issue, this study proposes a multimodal sensing approach that integrates visible–near infrared spectroscopy (VIS–NIR) and bioimpedance technology. By leveraging the complementary advantages of VIS–NIR spectroscopy technology and bioimpedance technology, this method combines bioimpedance analysis with VIS–NIR spectroscopy analysis to improve the accuracy of non‐destructive post‐harvest ripeness detection for kiwifruit. In this study, 120 “Hayward” kiwifruit samples stored at 10°C were used as the research objects. Their VIS–NIR spectral characteristics (450–860 nm) and bioimpedance properties (10 Hz–1 MHz) were measured, while physical and chemical indicators such as firmness, soluble solids content (SSC), and water loss rate were tested, and sensory evaluations were conducted. On this basis, a machine learning‐based ripeness grading model was developed. A three‐level data fusion strategy was adopted, including low‐level feature concatenation, intermediate‐level amplitude/phase combination, and high‐level consensus decision‐making. Additionally, the performance of BP (Back Propagation) neural networks, RBF (Radial Basis Function) neural networks, and SVM (Support Vector Machine) models was compared. The optimal grading performance was achieved by directly concatenating VIS–NIR spectral data and bioimpedance data and inputting them into the SVM model—yielding an accuracy of 87.39%, precision of 87.38%, and recall of 87.39%. This performance significantly outperformed single‐modality approaches (VIS–NIR spectroscopy: 77.62%; bioimpedance: 76.49%). This method provides a new technical pathway for non‐destructive quality detection and precise grading of post‐harvest kiwifruit.