在双波段可见/近红外光谱系统（350~1 100和1 000~2 500 nm）中, 由于两台仪器性能有所不同, 导致在波段重叠区域对同一样品测得的反射率不同, 出现数据交叉现象。 针对此问题, 提出一种波段连接数据融合的方法, 以期对两个波段的光谱进行更好的应用。 首先采集60个生鲜猪肉样品表面的反射光谱信息, 利用Savitzky-Golay（S-G）平滑和标准正态变量变换进行预处理, 然后利用单一波段和双波段光谱数据与猪肉品质参数（颜色参数L*, a*, b*, pH和蒸煮损失率）理化值建立偏最小二乘预测模型, 并分析比较。 利用提出的波段融合方法对两个波段重叠区域出现的交叉进行处理, 处理后的双波段光谱融合数据对参数L*, a*, b*, pH以及蒸煮损失率建模, 验证集的相关系数分别为0.948 8, 0.920 0, 0.950 5, 0.930 1和0.903 5, 模型效果与未融合前相当甚至更优。 采用无信息变量消除法方法进行特征变量筛选, 利用优选后的特征变量建立了更为简化的模型。 实验结果表明, 所提出的波段融合方法能够对两个波段光谱数据实现较好的融合, 利用融合后的光谱数据有利于建立更简化、 性能更佳的预测模型。

For dual band visible/near infrared spectroscopy system (350~1 100 and 1 000~2 500 nm), there exsits a band overlap and for the same sample the reflectivity data were unlike due to the performance difference between instruments. A band connection and data fusion method was proposed in this paper to make better use of the dual-band data. A dual-band visible/near-infrared spectroscopy system was built in the study to collect 60 pork samples’ reflectance spectra. The reflectance spectra of samples were performed with pretreatment methods of Savitzky-Golay (S-G) and standard normal variable transform to eliminate the spectral noise. Then partial least squares regression (PLSR) prediction models of pork quality attributes (color, pH and cooking loss) based on single-band spectrum and dual-band spectrum were established, respectively. For the cross of two band overlap, the data were connected and integrated using the method put forward in this paper and then PLSR models were established based on the integrated data. The PLSR model yielded prediction result with correlation coefficient of validation (Rp) of 0.948 8, 0.920 0, 0.950 5, 0.930 1 and 0.903 5 for L*, a*, b*, pH value and cooking loss, respectively. To simplify the model, uninformative variables elimination (UVE) was employed to select characteristic variables. The experimental results show that the proposed method was able to achieve a better fusion of the two band spectral data, and it was good for the establishment of a more simplified and better prediction model.