Non-destructive Detection Of Mold-damaged Chestnuts Based On Near-infrared Spectroscopy

Chestnuts have low storage stability and are prone to mildew after harvest; however, some mold-damaged chestnuts can not be recognized in appearance. The conventional method for rapid detection of mold-damaged chestnuts is floating, but it is time-consuming. Near-infrared spectroscopy is a kind of non-destructive detection method which classifies the samples through the analysis of the difference of spectra. In the present work, two ranges of spectra,600-1100nm and 1100-2500nm, respectively, were obtained by two near infrard spectrometers and then were classified based on PCA-DA (principle component analysis-discriminant analysis), SIMCA (soft independent modeling of class analogy) and LDA (linear discriminant analysis) methods. After choosing the characterization from the spectra, the characteristic information which has great contribution to the sample classification model was selected. Then, the model was established and validated. The effects of spectrum in different wavelength, pretreatment method and variation selection on the model were compared, and the results were as follows:(1) In PCA-DA model establishment, the best pretreatment method for 600-1100 nm and 1100-2500 nm spectra was "smoothing+MC (mean centering)+SNV (standard normal variate)". The accuracy rate of calibration set and validation set for 600-1100 nm spectrum were 99.46%and 94.76%, respectively. The accuracy rate of calibration set and validation set for 1100-2500 nm were 98.53%and 95.87%, respectively.(2) In SIMCA classification model establishment, the pretreatment method "first derivative + smoothing+detrending" was better than "first derivative+smoothing+SNV". The accuracy rate for 600-1100 nm spectrum was higher than that of 1100-2500nm.(3) In LDA model establishment, normalization was used as the pretreatment method for the spectrum of 1100-2500nm. The best wavelength for model establishment was 1100-1158 nm. Baseline method was used as the pretreatment method for the spectrum of 600-1100 nm and the best wavelength for model establishment was 800-1100nm. In general, the model established with the spectrum of 600-1100 nm was better than that of 1100-2500nm.