为了实现轻微损伤郎枣的快速无损检测，以完好和轻微损伤郎枣为研究对象，动态采集其可见/近红外光谱数据。依据光谱波段定义将采集的光谱数据分为可见光(Vis)、短波近红外(SW-NIR)、长波近红外(LW-NIR)、可见/短波近红外(Vis/SW-NIR)、近红外(NIR)和可见/近红外(Vis/NIR)等6个波段，分别选取各波段最佳预处理方法。采用连续投影法(SPA)和主成分分析法(PCA)分别对各波段光谱数据降维，以全波长、SPA提取的特征波长和PCA提取的主成分作为输入，分别建立偏最小二乘回归法(PLSR)和最小二乘支持向量机(LS-SVM)模型，通过比较预测集的判别准确率，确定最佳建模方法。结果表明，PLSR模型优于LS-SVM模型，SW-NIR波段较其余5个波段有更好的判别能力，所建SW-NIR-SNV-SPA-PLSR模型判别准确率为93.3%，为最佳模型。本实验为轻微损伤郎枣的快速无损检测和相关仪器的开发提供了理论基础。

For rapid and non-destructive detection of Lang jujubes with subtle bruises, dynamic visible/near-infrared (NIR) spectral data was collected to study intact and subtly bruised Lang jujubes. Based on the definition of different spectral ranges, the spectral data obtained were divided into six spectral ranges, namely, visible (Vis), short-wave NIR (SW-NIR), long-wave NIR (LW-NIR), Vis/SW-NIR, NIR, and Vis/NIR. The optimal pretreatment method for each range was selected. Successive projections algorithm (SPA) and principal component analysis (PCA) were used to reduce the dimensions of the full spectrum (FS). Using the characteristic wavelengths extracted by SPA, the principal component extracted by PCA and FS of the 6 spectral ranges as input variables, a partial least squares regression (PLSR) model and a least-squares support vector machines (LS-SVM) model were established. The optimal model was identified by comparing the discrimination accuracy of the prediction set. The results indicated that PLSR was more preferable than LS-SVM and the SW-NIR spectral range was optimal in comparison to the other five spectral ranges in terms of discriminatory power. The optimal model was identified as SW-NIR-SNV-SPA-PLSR and the discrimination accuracy of the prediction set was 93.3%. This study provide a reasonable theoretical basis for the discrimination of subtly bruised Lang jujubes and development of relevant instruments.

国家自然科学基金资助项目(31271973)；山西省自然科学基金资助项目(2012011030-3)