基于全局敏感性分析与机器学习的冬小麦叶面积指数估算

在小麦叶面积指数(leaf area index,LAI)的估算过程中,光谱变量与机器学习算法(MLs)相结合的方法具有较好的性能,但由于输入参数过多会导致数据冗余,使得计算效率降低。为了提高LAI估算的精度和MLs的计算效率,本研究提出了全局敏感性分析(global sensitivity analysis,GSA)与MLs相结合的方法(简称GSA-MLs)。首先,基于PROSAIL模拟数据集,利用GSA量化植被生长参数对Sentinel-2光谱变量的影响;此外利用4种变量筛选策略对所有光谱变量进行排序,并选择最优变量作为MLs的输入参数。然后,通过偏最小二乘回归(partial least square regression,PLSR)、支持向量机(support vector machine,SVM)和随机森林(random forest,RF)3种MLs对小麦叶面积指数(LAI)进行估算。结果表明:红边植被指数主要受叶绿素含量的影响,而短波红外相关的植被指数主要受等效水厚度的影响,所有光谱变量均会受到参数之间的交互作用。SLAI-SInteraction筛选得到的30个光谱变量在估算小麦LAI表现最佳(R²=0.94,RMSE=0.38)。并且在模型反演过程中运行时间缩短了54.13%。本研究提出了全局敏感性分析与机器学习相结合的方法,该方法提高了机器学习法估算LAI精度以及应用过程中的计算效率和机理性,该方法有较好的适用性。

Leaf area index estimation of winter wheat based on global sensitivity analysis and machine learning

In the estimation process of wheat leaf area index (LAI), the method combining spectral variables with machine learning algorithm (MLs) has better performance. However, too many input parameters will lead to data redundancy and reduce the computing efficiency. In order to improve the accuracy of LAI estimation and the efficiency of MLs calculation, a method combining global sensitivity analysis (GSA) and MLs(GSA-MLs) was proposed in this study. Firstly, based on the PROSAIL simulation dataset, GSA was used to quantify the effects of vegetation growth parameters on Sentinel-2 spectral variables. In addition, four variable screening strategies were used to sort all spectral variables, and the optimal variable was selected as the input parameter of MLs. And then, partial least square regression (PLSR), support vector machine (SVM) and random forest (RF), three MLs were used to estimate wheat leaf area index (LAI). The results showed that the red edge vegetation index was mainly affected by chlorophyll content, while the short-wave infrared vegetation index was mainly affected by equivalent water thickness. All spectral variables were subject to interaction between parameters. The 30 spectral variables screened by SLAI-SInteraction performed best in estimating wheat LAI (R²=0.94, RMSE=0.38). Moreover, the running time of model inversion was shortened by 54.13%. This study proposed a combination of global sensitivity analysis and machine learning. In addition to improving the accuracy of LAI estimation by machine learning method and the calculation efficiency in the application process, the machine theory in the application process of machine learning was improved and had good applicability.