【摘要】 为了克服传统植物凋落叶碳质量分数检测方法成本高、费时耗力等问题，探究基于可见光、近红外图像信息和机器学习算法构建凋落叶碳质量分数估测模型，为实时监测和快速获取植物凋落叶的碳质量分数提供技术手段。以白桦凋落叶为研究对象，利用可见光和近红外图片信息提取光学三原色（RGB）、六角锥体模型（HSV）和单色图像中的叶片颜色、纹理和形状特征共47个，然后利用主成分分析对特征变量进行降维，并利用遗传算法优化神经网络模型（GA-BPNN）、3种不同核函数的支持向量机回归模型（SVR）和随机森林回归模型（RFR）对叶片碳质量分数进行建模和预测。结果表明：在训练数据集上，RFR模型对叶片碳质量分数的拟合和预测效果最好(平均绝对误差(E_MA)=4.625 3,均方根误差(E_RSM)=5.608 7,平均百分比误差(E_MAP)=0.010 6,决定系数（R²）=0.834 8);在测试数据集上，GA-BPNN和RFR模型对叶片碳质量分数预测精度相似，RBF-SVR模型的拟合和预测效果最好(E_MA=6.529 2,E_RSM=7.925 2,E_MAP=0.015 0,R²=0.610 7),RBF-SVR模型与BPNN和RFR模型相比E_MA分别下降13.04%和13.27%、E_RSM分别下降8.6%和9.77%、E_MAP分别下降12.79%和13.29%。研究结果为快速无损获取凋落叶碳质量分数，及时预测凋落物的分解速率提供了新的方法。更多还原

【Abstract】 In order to overcome high cost and time consuming of traditional methods in detecting the carbon content of leaf litter, we explored the construction of a model for estimating the carbon content of leaf litter based on visible/near-infrared（NIR） image information and machine learning algorithms, and provided technical means for real-time monitoring and rapid acquisition of the carbon content of leaf litter. A total of 47 leaf color, texture and shape features in RGB, HSV and monochrome images were extracted using visible and NIR image information, then the feature variables were dimensionalized using principal component analysis, and the leaf carbon content was modeled and predicted using a genetic algorithm optimized neural network model（GA-BPNN）, a support vector machine regression model（SVR） with three different kernel functions, and a random forest regression model（RFR）. The results showed that the RFR model fitted and predicted the leaf carbon content best on the training data set(E_MA=4.625 3, E_RSM=5.608 7, E_MAP=0.010 6, and R²=0.834 8); on the test data set, the GA-BPNN and RFR models predicted the leaf carbon content with similar accuracy, the RBF-SVR model fitted and the RBF-SVR model had the best fit and prediction(E_MA=6.529 2, E_RSM=7.925 2, E_MAP=0.015 0, R²=0.610 7), and the RBF-SVR model decreased E_MA compared to the BPNN and RFR models by 13.04% and 13.27%, E_RSM decreased by 8.6% and 9.77%, E_MAP decreased by 12.79% and 13.29%, respectively. The study provided a new method for rapid and non-destructive acquisition of the carbon content of leaf litter and timely prediction of the decomposition rate of leaf litter.更多还原