基于光谱成像的猪肉新鲜度空间分布预测评价方法

Evaluation of Spectral Imaging-based Spatial Predictions of Freshness Spatial Distribution over Pork

新鲜度指标在像素位置缺乏微观参考值,因此将基于均值光谱的化学计量学模型应用到像素光谱时,无法对指标空间分布预测质量进行直接评价。提出了基于准度和精度的评价方法,以兴趣区域内各像素位置微观预测值的统计均值相对于理化检测值的决定系数和均方根误差作为准度评价指标;根据新鲜度指标的理论允许范围,以TVB-N微观预测值小于零的像素点在兴趣区域内所占比值作为精度评价指标。基于偏最小二乘回归,在可见近红外波段(550~970 nm),分别对全波段、利用连续投影算法精选的20个和6个特征波段建立新鲜度预测模型;采用5种不同带宽的光谱滤波,将滤波前后光谱所得指标空间分布预测结果进行比较。研究表明:经不同光谱预处理及化学计量学模型所得指标空间分布预测结果存在显著差异。尽管光谱均值滤波后像素光谱质量仍低于均值光谱,但指标空间分布预测准度恒等于预测模型本身;指标空间分布预测精度明显受到像素光谱质量及预测模型波段增益值的共同影响,前者占主导作用(R=0.72)。因此,本文的评价方法能够对基于光谱成像的化学计量学指标空间分布预测质量进行评价;利用线性化学计量学算法进行指标空间分布预测准度不会下降;在实践中,可以通过提高像素光谱信噪比和限制模型波段增益提高预测的精度。

Since the unavailability of local reference freshness values at individual pixels,a direct validation is impossible for the spatial distribution of pork freshness,i.e.,the freshness maps visualized through applying the chemometric models that are trained on average spectra of regions of interest(ROI)to the spectra at individual pixels within the ROIs.Therefore,a dual-criteria evaluation of the freshness maps that were produced through different chemometric systems coupled with varied spectral filtering on both accuracy and precision was proposed.The former was quantified by the coefficient of determination of prediction(R_(P)^(2))and the root mean square error of prediction between the chemical reference of ROIs and the average of the predictions at all individual pixels therein.The latter was quantified with the ratio of the pixels having negative TVB-N values to those of the ROI for a given subject,since the non-negativity according to the theoretical range of the freshness measurement.A bank of drastically different freshness maps of the same batch of pork were produced by using partial least squares regression(PLSR),over the visual/near infrared spectral range over 550~970 nm,both before and after spectral filtering using ideal average smoothing filters with five different bandwidths of 6 nm,18 nm,30 nm,42 nm,and 54 nm,respectively.The full range of consecutive wavebands,as well as 20 or 6 feature bands which were selected by successive projection algorithm(SPA),were used to form a collection of 18 combinations of bandwidth and the number of spectral bands to build chemometric models.Drastic difference resulted between the 18 approaches to visualization of freshness distribution.Analysis result showed,however,that all freshness maps were of good accuracy,equal to that of the chemometric models despite the lower quality of the spectra at individual pixels,even after spectral filtering,than those used in the training of models.And the precision of spatial predictions of freshness seemed to be co-determined by both spectra quality at individual pixels and the waveband-gains of chemometric models,and dominated by the former,R=0.72.It may be concluded that the spatial distributive predictions from imaging chemometrics can be objectively evaluated according to the statistics of the local predictions at pixels and the theoretic range of quality-indicating attributes;accuracy of quality-indicating maps,predicted on spectra at pixels,would not change from that of a linear chemometric system;better precision of spatial distribution prediction could be expected if spectral signal-to-noise ratio at pixels was improved and a chemometric model’s gains of wavebands were low.