Development of cold damage classifiers for peach flowers using hyperspectral imagery

초록

This study proposes a hyperspectral imaging-based analytical framework and artificial intelligence (AI) classification models for the early, non-destructive detection of cold injury in peach flowers during the blooming stage. Peach flowers are highly vulnerable to low-temperature stress, yet conventional diagnosis relying on visual inspection or fruit-stage assessment limits timely response and cultivar selection. To address this, hyperspectral images covering 400 - 1,000 nm were acquired, region-of-interest spectra were extracted, and standardized preprocessing methods such as interquartile range (IQR) and standard normal variate (SNV) were applied. Machine learning and deep learning classification models were developed and evaluated under different temperature conditions. Among the tested configurations, the IQR - convolutional neural network (CNN) combination achieved the highest accuracy and reproducibility, indicating effective learning of subtle spectral responses to cold stress. Mean spectral analysis showed decreased reflectance below 3℃, likely associated with cell structural weakening and moisture reduction. Sensitivity analysis revealed notable responsiveness at 400 - 420 nm, 700 - 800 nm, and above 900 nm, and model-based feature importance highlighted 520 - 525 nm, 575 - 580 nm, 675 - 680 nm, 785 - 790 nm, 870 - 880 nm, 890 - 920 nm, and near 940 nm as discriminative regions. Among these, 675 - 680 nm, 785 - 790 nm, 870 - 880 nm, and 890 - 920 nm were interpreted as physiologically relevant to cold-injury responses. Overall, this study addresses limitations of fruit-stage assessments by presenting a blooming-stage, non-destructive detection framework with potential applications in cold-injury early warning, cold-tolerant cultivar development, and orchard management decision support.

키워드