Algorithms Research of near Infrared Spectral Database of Apples

Abstract

A near infrared spectral database system of apples was studied. A new spectral matching algorithm based on Jaccard similarity coefficient (SMA-JSC) is proposed for higher spectral matching accuracy. A total of 840 apples of seven different varieties were used to evaluate the performance of this method. At the same time, a comparison was done among the existing spectral matching algorithms with spectral peak information (SMA-P) and the existing spectral matching algorithms with full spectra (SMA-FS). The highest accuracies of these existing SMA-P and SMA-FS were 48.9% and 72.57%, both of which were quite low mainly because of noise. For SMA-JSC, the first-order derivative was calculated and transformed into binary values (with only 0 or 1) to eliminate the influence of noise. The accuracy of our proposed algorithm was 94.1% for the calibration samples and 94.3% for the validation samples. Also, using linear discriminate analysis (LDA), we compared the existing SMA-P, the existing SMA-FS and our new proposed algorithm. The best result obtained using LDA was 88.0% (with the raw spectra) which was larger than the accuracies of the existing SMA-P and SMA-FS but less than the accuracy of our proposed SMA-JSC. In addition, several important spectral data parameters were optimised for automatic spectral peak detection without manually setting parameters (optimised parameters: spectral resolution was 32 cm−1, threshold of peak width was 29 spectral data points and threshold of peak shape index was 0.005). A selective spectral smoothing algorithm is also proposed to protect peak bands. With these methods, spectral peaks could be detected 100% correctly without manually setting parameters.