植物凋落物是联结生物有机体合成和分解的桥梁, 通过物质流、 能量流及信息流深刻地影响了陆地生态系统的结构、 功能及关键生态过程。 自然生态系统中地表凋落物通常以混合物的形式分解, 尤其是在物种极其丰富的亚热带常绿阔叶林中。 受样地内树种组成影响, 叶凋落物往往属、 种混杂, 非专业人士难以实现准确鉴别, 这为后续凋落物分解研究带来一定的困难。 近红外光谱分析技术作为一种快速无损的检测手段, 已经成功应用于牛肝菌、 柑橘、 水稻等的种类鉴别。 该技术为解决叶凋落物鉴定这一难题提供了新的途径。 该研究收集云南哀牢山典型中山湿性常绿阔叶林6种优势树种叶凋落物共计540份, 获取样品近红外漫反射光谱, 分析不同种类叶凋落物平均光谱图特征。 建模时, 使用Kennard-Stone算法将540个样品数据以2∶1比例分为训练集与验证集, 其中360个样品数据用于叶凋落物分类模型的建立, 180个样品数据用于叶凋落物分类模型的验证。 使用标准正态变量变换(SNV)、 Savitzky-Golay卷积平滑(SG)、 多元散射校正(MSC)、 导数处理(Derivative)等单一与组合算法对光谱数据进行预处理, 并采用主成分分析(PCA)与正交偏最小二乘判别分析(OPLS-DA)2种模式识别方法对不同树种叶凋落物进行鉴别。 结果表明: (1) 叶凋落物近红外光谱主成分分析得分图中各组数据呈现交织状态, 虽然经SNV+SG方法预处理后, 光谱数据得到优化, 变色锥与舟柄茶与其他树种出现一定的区分, 但仍然无法实现6种叶凋落物的准确鉴别。 (2) SNV+SD预处理方法结合OPLS-DA建立的模式识别模型, 效果最好, 因变量累计拟合指数为0.922, 模型累计预测能力指数为0.894, 置换检验显示模型未过度拟合, 训练集与验证集识别率均为100%。 研究表明, 在对样本近红外光谱进行预处理优化的基础上, 结合有监督的OPLS-DA模式识别方法, 可以实现不同树种叶凋落物的准确鉴别, 为后续植物凋落物研究提供了有力的技术支撑。

As a bridge between the synthesis and decomposition of a biological organisms, plant litter impacts the structure, function and key ecological processes of terrestrial ecosystems through material, energy and information flow. Litters decompose as species mixtures in natural systems, especially in species-rich subtropical evergreen forests. It is difficult to accurately identify leaf litter for non-professionals due to complex tree species in the field. Besides, misidentifications cause many problems for thesubsequent litter decomposition research. As a fast and nondestructive analysis method, near-infrared spectroscopy has been successfully applied to identify boletus, citrus and rice. The technique mentioned above systems provided a new way to solve problems of leaf litter identification. In this study, 540 leaf litter samples of 6 dominant tree species of typical mid-mountain moist evergreen broad-leaved forests in the Mts. Ailaoshan were collected. The diffuse reflectance spectra were recorded on individual samples using an Antaris Ⅱ FT-NIR analyzer and the average spectral characteristics of different litter species were analyzed. During each modeling, 540 sample data were divided in to the training set and test set at a ratio of 2∶1 by using the Kennard-Stone algorithm. 360 sample data were used to develop discriminant models and 180 sample data were used to test the models. Single and combined spectral pretreatment methods (SNV, SG, MSC, and Derivative) were applied to improve the performance of discrimination models. Two qualitative pattern recognition methods (i. e., principal component analysis, PCA and orthogonal partial least-squares discrimination analysis, OPLS-DA) were conducted to identify the species of leaf litter. The results showed that: (1) the spectra data of different litter groups intertwined in the PCA score plot. Using SNV+SG as the pretreatment of spectra could improve the model parameter. PCA method cannot identify the leaf litter of six tree species, though Castanopsis wattii and Hartia sinensis can be separated from the rest litter species using the improved discriminant model. (2) SNV+SD pretreatment method combined with the OPLS-DA algorithm was used to develop discriminant models and showed excellent prediction ability (training set=100%; validation set=100%). Key statistical parameters of this model including R2Ycum and Q2Cum were 0.922 and 0.894, respectively. The permutation test indicated that the discriminant model was not overfitted. Our study indicated that NIR calibration models built with OPLS-DA algorithm have a good discriminative ability for different leaf litter species, and thus provide definite technological support for further plant litter research.