基于近红外光谱技术的茶鲜叶海拔高度判别模型建立
Establishment of discrimination model for different elevation fresh tea leaves based on near infrared spectroscopy

以不同海拔高度的茶鲜叶为研究对象，扫描获取其近红外光谱（NIRS）并筛选特征光谱区间后，分别应用逐步多元线性回归法（SMLR）、主成分回归法（PCR）和联合区间偏最小二乘法（Si PLS）建立茶鲜叶海拔高度预测模型。结果表明，在5 542.41～6 888.48 cm-1区间内，对原始光谱进行一阶导数＋3点Norris平滑预处理后，建立的SMLR模型预测集相关系数和预测均方差分别为0.800 5和0.486；在4 929.16～6 965.62 cm-1区间内，当主成分数为3时，对原始光谱进行一阶导数＋3点Norris平滑预处理后，建立的PCR模型预测集相关系数和预测均方差分别为0.803 6和0.472；当将光谱划分为18个子区间、因子数为13时，选用\[5 8 11 17\]4个子区间建立的Si PLS模型预测集相关系数和预测均方差分别为0.944 3和0.295。经比较， Si PLS模型预测结果最佳。
There is a certain relationship between the quality of fresh tea leaves and the elevation of growing, but at present, it is no effective method to discriminate the elevation of fresh leaves picked. In this study, fresh tea leaves of different elevation were used as the research objects, after near infrared spectroscopy scanned and the characteristic spectral interval selected, the prediction models of elevation of fresh tea leaves were established by stepwise multiple linear regression (SMLR), principal component regression (PCR) and synergy interval partial least squares (Si PLS). The results showed that, the correlation coefficient and root mean square error of prediction set was respectively 0.800 5 and 0.486 by SMLR method, which used the spectroscopy in the range of 5 542.41-6 888.48 cm-1 and the first derivative +3 point Norris smoothing pretreatment; the correlation coefficient and root mean square error of prediction set was respectively 0.803 6 and 0.472 by PCR method, which used the spectroscopy in the range of 4 929.16-6 965.62 cm-1 and the first derivative + 3 point Norris smoothing pretreatment; the correlation coefficient and root mean square error of prediction set was respectively 0.944 3 and 0.295 by Si PLS method, which contained 18 spectral intervals combined with \[5 8 11 17\] of four subintervals and 13 factors. By comparison, the Si PLS model has the best prediction results. It was preliminary realized to discriminate the elevation of fresh tea leaf samples rapidly and nondestructively by using NIRSSiPLS method.