Structured hyperspectral imaging and machine learning for non-destructive kiwifruit firmness prediction, classification, and intelligent post-harvest management

Abstract

Kiwifruit firmness is a critical quality attribute influencing consumer acceptance and post-harvest logistics. However, the unpredictable ripening process creates challenges in quality assessment and purchase decisions of consumers. This study proposes a structured hyperspectral imaging (S-HSI) system integrated with machine learning algorithms to enable non-destructive prediction and classification of kiwifruit firmness during shelf-life. A Random Forest (RF) regression model was developed based on spectral data acquired from S-HSI, achieving R^²_c with 0.8697 and R^²_p with 0.8204, surpassing the conventional hyperspectral imaging (HSI) method by 3.32 % and 7.98 %, respectively. According to the changes of firmness during shelf-life, kiwifruits were divided into two categories for guiding the purchase behavior: ready-to-eat (<20 N) and requiring storage (>20 N). The artificial neural network (ANN) classifier trained on S-HSI spectral features achieved 96.04 % classification accuracy, outperforming HSI-based models. Shapley Additive exPlanations analysis identified critical spectral regions in the near-infrared range, highlighting the enhanced feature extraction capability of S-HSI. Furthermore, Pearson correlation analysis between microscopic hyperspectral reflectance and S-HSI/HSI confirmed a stronger correlation for S-HSI, explaining its superior predictive accuracy. This research demonstrates the potential of S-HSI in food engineering applications, enabling real-time, high-accuracy firmness assessment for intelligent post-harvest quality control and smart supply chain management.