基于高光谱成像的牧草粗蛋白含量检测研究

粗蛋白（CP）是评价牧草营养价值和品质参数的关键指标。快速、准确地对牧草中粗蛋白含量进行评估在畜牧业生产研究中具有重要意义。为确定牧草粗蛋白含量的高光谱特征波段及最优检测模型，研究分别于2017年5月至9月间在黑龙江省杜尔伯特自治区的人工牧草场（羊草）内每月随机选取35个样本，5个月共采集175个样本。采样时在样本点处放置1 m×1 m的样方，将样方内所有牧草全部齐地面收割采集后称重并冷藏保存。将样本带回实验室后，立即进行牧草叶片高光谱图像采集，同时采用凯氏定氮法对采集的牧草样本进行粗蛋白化学值测定，以此建立牧草粗蛋白含量高光谱数据集。研究首先通过Savitzky-Golay卷积平滑（SG）、多元散射校正（MSC）、变量标准化（SNV）、一阶导数（1-Der）和直接正交信号校正（DOSC）方法5种预处理方法对高光谱数据进行处理后分别建立偏最小二乘回归（PLSR）检测模型，从中确定最优预处理方法。利用最优预处理结果，分别采用连续投影算法（SPA）和随机蛙跳算法（RF）进行牧草粗蛋白含量的特征波段选择，并利用选择结果分别进一步建立PLSR模型，以此确定适合粗蛋白含量的特征波段选择方法，确定最优高光谱检测模型。结果表明，在五种高光谱预处理方法中，基于SNV方法预处理后所建立的高光谱PLSR模型表现最优（R2-P=0.929，RMSE-P=6.344 mg·g-1，RPD=4.204）。利用连续投影算法筛选的粗蛋白含量特征波长为30个，分布于530～700和940～1 000 nm范围内。经随机蛙跳算法确定的粗蛋白含量特征波段为6个，分别为826.544，827.285，828.766，971.012，972.494和973.235 nm。因此，该研究中牧草粗蛋白含量最优高光谱检测模型为SNV-RF-PLSR（R2-P=0.933，RMSE-P=6.034 mg·g-1，RPD=4.322），模型精度较高。该研究结果为牧草粗蛋白含量的高光谱检测提供了最优模型和理论基础，同时为指导草业生产开拓了新的技术思路。

Research on Crude Protein of Pasture Based on Hyperspectral Imaging

Crude protein (CP) is the key parameter for evaluating nutritive value and quality of pasture. It has a great significance for evaluating crude protein content of pasture quickly and accurately in animal husbandry. For confirming the hyperspectral characteristic bands and optimal detection model of crude protein content in pasture, we randomly selected thirty-five sample plots each month from May to September, 2017 in Dorbet, Heilongjiang Province, one hundred and seventy-five samples for all. A 1 m×1 m quadrangle was placed at the sample point during sampling, and all the aboveground pastures in the quadrangle were collected, weighed and stored in cold storage. After carrying the samples the laboratory, we collected the hyperspectral information immediately and determined the chemical values of crude protein by Kjeldahl determination, establishing the hyperspectral dataset of crude protein content. We used five pre-processing methods including SG, MSC, SNV, 1-Der, DOSC to process the hyperspectral data and then, built the PLSR models for confirming the optimal pre-processing method. Based on the optimal pre-processing result, the characteristic bands of crude protein were selected by successive projections algorithm and random frog algorithm, then the PLSR models were built for confirming the optimal selection method of characteristic variables and the optimal hyperspectral detection model. The results showed that the hyperspectral detection model based on SNV was the best in the five pre-processing methods. Thirty bands were selected by SPA and distributed in 530 to 700 nm and 940 to 1 000 nm. Six bands were selected by RF, and respectively were 826.544, 827.285, 828.766, 971.012, 972.494 and 973.235 nm. Therefore, the optimal hyperspectral detection model was SNV-RF-PLSR in this research, and the accuracy of model was good. The results of this research provided an optimal model and theoretical basis for hyperspectral detection of crude protein in pastures and in addition, developed new technique solutions for guiding the production of grassland industry.