Comparison of mutton freshness grade discrimination based on hyperspectral imaging, near infrared spectroscopy and their fusion information

Abstract

The discrimination models of mutton freshness grades based on hyperspectral imaging (HSI, 400-1,000 nm), near infrared spectroscopy (NIRS, 900-2,500 nm) and their fusion information were established and compared in this study. Mutton freshness was divided into three grades by the comprehensive evaluation criterion. Then the characteristic variables were screened from preprocessed full-band variables, and discriminant models were established and compared based on the full-band variables, the characteristic variables and fusion information. The results showed that both full-band models of HSI and NIRS showed good performance, and the accuracies of all datasets were higher than 92.6%. The model accuracy decreased gradually with the decrease of the variables number when modeling with characteristic variables. When modeling with fusion information, the extreme learning machine models had better performance, and the accuracies of all datasets were over 92.7%. Comprehensively considering the accuracy, the cost, the efficiency of data acquisition and analysis, the fusion of HSI and NIRS had no significant effect compared to the single sensor with full-band variables, and NIRS was possibly more economical and practical than HSI. This study provides reference and theoretical basis for selection of sensors in future studies and development of simplified, low-cost detection devices for mutton freshness grade. Practical Applications Based on the full-band variables, the characteristic variables, and fusion information of HSI (400-1,000 nm) and NIRS (900-2,500 nm), the discrimination models of mutton freshness grades were established and compared. The characteristic variables were screened by the methods of CARS, GA, SPA and their synergic methods. We could come to a conclusion that both full-band models of HSI and NIRS showed good results. Comprehensively considering the accuracy, the cost, the efficiency of data acquisition and analysis, the fusion of HSI and NIRS had no significant effect compared to the single sensor based on the full-band variables, and NIRS was possibly more economical and practical than HSI. This study provides reference and theoretical basis for selection of sensors in future studies and development of simplified, low-cost detection devices for mutton freshness grade.

Quantitative study on impact damage of honey peaches based on reflection, absorbance, and Kubelka‐Munk spectrum combined with color characteristics

“…There are some invalid or interfering information in the spectra, which can reduce the prediction accuracy of the model, and the amount of data contained in the spectra makes the running time of the model increase. Using the characteristic wavelength extraction algorithm to select the most representative characteristic wavelengths from the spectra can effectively improve the efficiency of the model operation 24–26 . In this study, the CARS algorithm was used to select the characteristic wavelengths.…”

“…Using the characteristic wavelength extraction algorithm to select the most representative characteristic wavelengths from the spectra can effectively improve the efficiency of the model operation. [24][25][26] In this study, the CARS algorithm was used to select the characteristic wavelengths.…”

Application of portable near‐infrared spectroscopy technology for grade identification of Panax notoginseng slices

“…Extreme learning machine (ELM) (Bian, Li, Fan, Guo, & Chang, 2016) is a very popular machine learning algorithm, which used to train single hidden layer feedforward neural network (SLFN). This study used default function, default parameters, and the number of neurons in the hidden layer as 20 for training (Zhu et al, 2021).…”

Research Progress on Quality Detection of Livestock and Poultry Meat Based on Machine Vision, Hyperspectral and Multi-Source Information Fusion Technologies

“…Experimental results demonstrated that the model achieved prediction accuracies of 0.91 and 0.94 for pork and chicken meat, respectively. Zhu et al [54] conducted a study on lamb freshness using hyperspectral imaging and near-infrared spectroscopy imaging. They collected freshness data for three different grades of lamb meat according to comprehensive evaluation standards for meat quality.…”