结合近红外高光谱成像技术和不同的判别分析模型对4种国产咖啡豆品种进行了快速无损判别。通过高光谱成像仪提取874~1 734 nm波段内的光谱数据, 去除首尾噪声波段后, 分别基于925~1 680 nm波段的全谱波段和通过连续投影算法(SPA)选择的特征波长, 建立了偏最小二乘判别分析(PLS-DA)、随机森林(RF)、K最邻近算法(KNN)、支持向量机(SVM)模型和极限学习机(ELM)5种判别分析模型。 基于上述判别模型对咖啡豆品种进行鉴别; 然后通过准确率、命中率和否定率3个参数对鉴别结果进行了评价。实验显示, 基于全谱和特征波段建立的模型均取得了较好的判别效果, 其中ELM模型效果均为最优, 每个品种建模集和预测集的准确率、命中率和否定率均在93.5%以上。研究结果表明, 基于近红外高光谱成像技术结合模型判别分析方法可以实现对国产咖啡豆品种的识别, 特征波长的选择减少了变量数, 但判别效果与全谱相当。

Four different Chinese domestic coffee beans were identified rapidly by combining near infrared hyperspectral imaging technique and five kinds of discriminant models . A near-infrared hyperspectral imaging system covering the spectral range of 874-1 734 nm was set up to capture hyperspectral images of coffee bean samples. The head and end of the spectra with obvious noises were removed, and the spectral data in the range of 925-1 680 nm were extracted to establish discriminant models in the experiment. The sensitive wavelengths were selected from the full spectra by Successive Projections Algorithm (SPA). Five discriminant methods, including Partial Least Square-discriminant Analysis (PLS-DA), Random Forest (RF), K-nearest Neighbor algorithm (KNN), Support Vector Machine (SVM) and Extreme Learning Machine (ELM) were applied to the establishment of discriminant models based on the full spectra and the selected sensitive wavelength variables. The properties of the models were compared and valuated by three parameters, sensitivity, precision and specificity. Among all discriminant models, the ELM models based on the full spectra and the selected sensitive wavelength variables show the best identification results, respectively. For each coffee bean cultivar, the sensitivity, precision and specificity of ELM models based on full spectra and the sensitive wavelengths are all over 93.5% in both the calibration set and the prediction set. It concludes that Chinese domestic coffee beans could be identified by near-infrared hyperspectral imaging combined with discriminant models rapidly. Selecting the sensitive wavelengths reduces variables, but the identification effect is the same as that of the full spectra.