An Auditable Input‐Governance Layer for Online Line‐Scan Hyperspectral Inspection of Kiwifruit: A Case‐Study Evaluation

ABSTRACT This study aimed to improve the reliability and traceability of online line‐scan hyperspectral inspection of kiwifruit by evaluating an auditable input‐governance layer before downstream quality prediction. Candidate regions of interest (ROIs) were extracted from 800.34 nm reflectance images using column masking, Otsu thresholding, morphological processing, and connected‐component labeling. The governance layer applied configuration‐specific false‐target removal, geometric and spectral validity checks, review prioritization based on principal component analysis and robust Mahalanobis distance, and batch‐level input monitoring. A fruit‐level closed‐loop dataset from five labeled acquisition batches included 124 fruits with soluble solids content (SSC) measurements and 96 fruits with firmness measurements. Three prespecified regressors—partial least squares regression, support vector regression, and random forest regression—were used to compare upstream policies, with performance reported as mean ± standard deviation, ratio of performance to deviation, and paired Wilcoxon tests. The complete policy retained 122 of 124 SSC‐labeled fruits and showed a consistent but modest improvement trend for SSC prediction. Under repeated random splits, SSC root‐mean‐square error decreased from 1.494 to 1.432 °Brix for support vector regression and from 1.523 to 1.429 °Brix for random forest regression. Under leave‐one‐batch‐out evaluation, the complete policy gave the lowest observed SSC error across all three regressors, but it was not statistically superior to the strip‐removal baseline in most comparisons, and firmness results were inconsistent. The layer is therefore best viewed as a case‐study demonstration of auditable input‐side quality governance for the current line‐scan setup.