【摘要】 为研究检测红松籽仁蛋白质含量的近红外光谱分析技术，在用变量标准化校正+一阶导数+小波变换对原始光谱进行预处理的基础上，分别运用主成分分析、改进型局部线性嵌入、局部切空间对齐、黑塞特征映射进行光谱数据的降维处理，分别构建偏最小二乘、岭回归、支持向量回归、极度梯度提升数学模型。结果表明，改进型局部线性嵌入+支持向量回归法建立的参数优化模型质量最佳。其降维方法优化参数为：维度取4,邻域数取50;验证集均方差均值为0.568 1,验证集皮尔逊相关系数均值达0.940 8。可见，模型的预测结果是可靠的，能够实现对红松籽仁蛋白质含量的无损、准确检测。更多还原

【Abstract】 To study the near-infrared spectroscopy for protein content detection in Korean pine seed kernels, principal components analysis（PCA）, modified locally linear embedding（MLLE）, local tangent space alignment（LTSA） and Hessian based locally linear embedding（HLLE） were used separately to reduce dimensions of the spectroscopic data, based on pretreatment of the original spectrum by standard normalized variate（SNV）+first derivative(1^st-Der)+Symlet4(SNV+1^st-Der+Sym4) method. Partial least square（PLS）, ridge regression（Ridge）, support vector regression（SVR） and extreme gradient boosting（XGBoost） were adopted separately to establish mathematical models. The results showed that, the quality of the parameter optimization model established by MLLE+SVR method was the best. The optimized parameters for dimension reducing were as follows: the dimension（n-components） was four, the neighborhood number（n-neighbors） was 50, the mean value of mean squared error of validation（mean-MSEV） was 0.568 1, and the mean value of Pearson correlation coefficient of validation（mean-PCCV） was 0.940 8. Therefore, the prediction results of the model is reliable, and non-destructive, accurate and quantitative detection of protein in Korean pine seed kernels can be realized.更多还原