Hyperspectral band selection and band combination has become a powerful tool and have gained enormous interest among researchers. An important task in hyperspectral data processing is to reduce the redundancy of the spectral and spatial information without losing any valuable details that are needed for the subsequent detection, discrimination and classification processes. An integrated principal component analysis (PCA) and Fisher linear discriminant (FLD) method has been developed for feature band selection, and other pattern recognition technologies have been applied and compared with the developed method. The results on different types of defects from cucumber and apple samples show that the integrated PCA-FLD method outperforms PCA, FLD and canonical discriminant methods when they are used separately for classification. The integrated method adds a new tool for the multivariate analysis of hyperspectral images and can be extended to other hyperspectral imaging applications.

Dimensionality reduction not only serves as the first step of data processing that leads to a significant decrease in computational complexity in the successive procedures, but also a research tool for determining optimal spectra requirement for online automatic inspection of fruit. In this study, the hyperspectral research shows that the near infrared spectrum at 753nm is best for detecting apple defect. When applied for online apple defect inspection, over 98% of good apple detection rate is achieved. However, commercially available apple sorting and inspection machines cannot effectively solve the stem-calyx problems involved in automatic apple defects detection. In this study, a dual-spectrum NIR/MIR sensing method is applied. This technique can effectively distinguish true defects from stems and calyxes, which leads to a potential solution of the problem. The results of this study will advance the technology in fruit safety and quality inspection and improve the cost-effectiveness of fruit packing processes.