Research On The Establishment And Transfer Method Of The Spectral Detection Model For The Freshness Of Mandarin Fish From Different Origins

Mandarin fish is tasty and rich in nutrients,and is one of the most popular freshwater fishes.The most important indicator of the quality of mandarin fish is freshness.The spectral detection model can realize rapid and non-destructive detection of fish meat,but this model is only suitable for samples measured by the same instrument and from the same origin.For the prediction of samples measured by different instruments and from different origins,poor model prediction or model unavailability can occur.Therefore,it is of great significance to carry out model transfer research and establish a fish freshness detection model with high applicability and good stability.In this article,based on near-infrared spectroscopy and hyperspectral technology,and taking the TVB-N content of mandarin fish samples as the freshness evaluation index,a freshness detection model of mandarin fish samples from three origins in Yangxin,Jiayu and Guangzhou was established,cross-regional detection of mandarin fish samples using a model transfer method.The main findings are as follows:（1）Based on near-infrared spectroscopy technology,combined with six different preprocessing methods,the optimal detection models for the full wavelength of mandarin fish freshness in three origins were determined.Among them,the full-wavelength model established by the original spectrum of the Yangxin origin sample has the best effect.The model training set correlation coefficient R_c=0.9884,the root mean square error RMSEC=0.5235mg N·100g^-1,the prediction set correlation coefficient R_p=0.9389,and the root mean square error RMSEP=1.2059mg N·100g^-1;the correlation coefficients of the simplified model prediction sets established by combining the three characteristic wavelength algorithms are all greater than 0.79,the simplified model established by the Yangxin origin sample combined with the UVE algorithm has the best effect,and the model training set correlation coefficient R_c=0.9091,Root mean square error RMSEC=1.4429mg N·100g^-1,prediction set correlation coefficient R_p=0.9001,root mean square error RMSEP=1.5275mg N·100g^-1.（2）Based on hyperspectral technology,the optimal full-wavelength model and simplified model were determined for the freshness detection of mandarin fish in three origins.The correlation coefficients of the prediction sets of the optimal full-wavelength models established by the mandarin fish samples from the three origins combined with the six preprocessing methods are all greater than 0.91.Among them,the full-wavelength model established by the original spectra of the mandarin fish samples from the Guangzhou origin has the best effect,and the model training set correlation coefficient R_c=0.9505,root mean square error RMSEC=1.2187mg N·100g^-1,prediction set correlation coefficient R_p=0.9401,root mean square error RMSEP=1.1972mg N·100g^-1;simplified model prediction set established by combining three characteristic wavelength algorithms The correlation coefficients are all greater than 0.82,and the simplified model established by the Yangxin origin sample combined with the CARS algorithm has the best effect,Model training set correlation coefficient R_c=0.9215,root mean square error RMSEC=1.0958mg N·100g^-1,prediction set correlation coefficient R_p=0.9314,root mean square error RMSEP=0.8675mg N·100g^-1.The establishment of near-infrared and hyperspectral full-wavelength models and simplified models lays the foundation for model transfer studies in the full-wavelength and characteristic wavelength ranges.（3）The differences of the mandarin fish samples from the three origins were compared and analyzed by the average spectrum method,the principal component score spatial distribution method and the model validation method.Combining three model transfer methods,the near-infrared spectrum and hyperspectral model transfer are realized respectively.The near-infrared spectrum model combined with the DS algorithm model has the best transfer effect.When the standard sample number is 20 and 21,the main model is used for Jiayu and Guangzhou origin.The sample prediction correlation coefficients are 0.7852 and 0.7536,and the model performance is improved by 101%and179%respectively;the hyperspectral model combined with the DS-CARS algorithm model has the best transfer effect.The predicted correlation coefficients of samples from Guangzhou origin are 0.7612 and 0.7428,and the model performance is improved by 144%and 287%respectively.The combination of CARS and DS algorithm can effectively reduce the number of standard samples selected and greatly reduces the complexity of operations.The results of this study show that near-infrared spectroscopy and hyperspectral technology are feasible for the freshness detection of mandarin fish.