Research On Nondestructive Determination Of Cocoon Quality Based On Spectroscopy And Hyperspectral Imaging Techniques

As the traditional economic agriculture in China, sericulture has been playing an important historical role in export and social economy development with much attention and support. Cocoon was the most important product in sericulture, which could be traded in the market as the raw materials for silk production. So the the precise evaluation of the cocoon quality was useful for pricing and grading. However, with the development of market economy and scientific and technological progress, the current cocoon quality evaluation method in China has turned to be inopportune with complicated operation, low implementation and poor technical level. As the the world’s largest cocoon producer and exporter, what we need is not only just the absolute advantage in quantity, but also the effective competitiveness of cocoon quality in the international market. In order to maintain the international status of our sericulture production, it is necessary to research the cocoon quality detection methods, to improve the informationalized level of cocoon production and management, to promote the further development of sericulture industry in China. This study is mainy focused on the nondestructive detection research of cocoon quality, according to the current evaluation criterions and the problems exsisted in real. The main research work and achievement were as follows:(1) The mounting time of fresh and dried cocoon were detected with the visible and near infrared spectroscopy between400～1000nm. The spectral data was pre-processed with baseline correction, and the effective wavelengths were selected by elimination of uninformative variables (UVE) and successive projections algorithm (SPA). The Bayes discrimation models based on the effective wavelengths were built with the prediction accuracy of91.11% and75.56% for fresh and dried cocoon mounting time detection, separately.(2) The regression coefficient (RC) and competitive adaptive reweighted sampling (CARS) were used for effective wavelengths selection between1250～2500nm. And the Bayes discrimation models based on the effective wavelengths were built with the prediction accuracy of70.00% and47.77% for fresh and dried cocoon mounting time detection, separately. The results showed that the effect for cocoon mounting time detection of near infrared spectroscopy was worse than visible and near infrared spectroscopy. (3) The rapid detection models for moisture content (MC) of fresh cocoon layer and dry weight (DW) of the cocoon layer were built. The wavelengths selection results of RC, UVE and CARS were compared, and secondly selection with SPA was promoted for better effective wavelengths selection results. The correlation coefficients of multiple linear regression (MLR) models for MC and DW prediction (Rp) were0.8473and0.8143, separately.(4) According to the spectral characters of water contained in cocoon layer between1250～2500nm, the MLR models for MC and DW detection of fresh cocoon were built with near infrared spectroscopy. The effects of5methods for spectral data pre-proccsing were researched. And the wavelength selection results of RC-SPA, UVE-SPA and CARS-SPA were systematically compared with the correlation coefficients and root mean square errors of PLS models built based on the selected wavelengths. Finally,9wavelengths were selected for MC detection by RC-SPA, and6wavelengths were selected for DW detection by CARS-SPA. The Rp of MLR models were0.7873å'Œ0.6992.(5) The sampling methods of calibration set, as random, concentration sort, Kennard-Stone, sample set partitioning based on joint x-y distances (SPXY) were all used for MC and DW detection. And the sampling results showed that the samples selected by Kennard-Stone were better than others in representativeness of all samples.(6) The results for sericin dissolubility (SD) of dried cocoon layer detection based on400-1000nm and1250～2500nm spectral data were compared.13effective wavelengths were selected from4000～1000nm by CARS-SPA, and9effective wavelengths were selected from1250～2500nm by UVE-SPA. The Rp of least square support vector machine (LS-SVM) models built based on the selected wavelengths were0.8709and0.6893. The results showed that the effect for cocoon SD detection of near infrared spectroscopy was worse than visible and near infrared spectroscopy.(7) The hyperspectral imaging system was applied for the detection of cocoon quality. The Rp of models based on the hyperspectral data between450～900nm were0.4969for MC detection,0.7838for DW detection, and0.5585for SD detection. And the prediction accuracy of models for mounting time of fresh and dried cocoon were73.33% and64.44%, separately.(8) The hyperspectral images of healthy and unhealthy cocoons were collected and processed by digital picture processing technique. The area-of-interest images of yellow spotted cocoons, cocoons pressed by cocooning frame and malformed cocoons were achived. And the prediction accuracy of Bayes discrimination model based on characteristic parameters of area-of-interest images was72.50%.