基于植被冠层光谱BRDF模型的LAI反演研究

发布时间：2019-06-02 15:32

【摘要】：叶面积指数(Leaf Area Index,LAI)是指单位水平地表上方所有叶片面积的总和,是衡量生物圈与大气圈能量中物质循环和能量流通、植被生长态势和监控全球气候变化的一个重要指标。传统直接测量方法虽然在精度上能够满足实际应用的需要,但该方法具有很大的破坏性且人力物力的投入过大,因而对于大面积LAI数据获取还存在很大难度;遥感监测属于非接触间接测量,以其能够大面积、高效、全天候、非破坏性测量等优势被广泛应用。在遥感影像数据获取的过程中,地表反射率往往作为反演地物理化及生物特性的重要参量而被广泛使用,但高精度地表反射率产品通常与传感器成像几何姿态以及大气环境密切相关。本文研究和讨论了太阳天顶角和传感器成像姿态对植被光谱的影响,提出一种基于植被冠层BRDF效应的LAI反演方法,该方法应用PROSAIL模型获取模拟数据,并将此作为样本集输入到BP神经网络(Back Propagation Artificial Neural Network,BP-ANN)进行训练构建反演模型。文章主要工作和相应结论如下:(1)、根据实测数据、参考资料及先验知识,构建植被辐射传输模型多角度参数数据库,在参数数据库基础之上模拟植被冠层光谱数据,根据模拟数据角度信息,计算半经验Ross-Li(Ross thick-Li transit)模型核参数。(2)、设计RVI、DVI、NDVI、MSR、EVI1、EVI2、RDVI、ARVI、SAVI、OSAVI、MSAVI、NLI 12个植被指数,分别与LAI建立经验统计模型。利用模拟验证样本和实测小麦数据对各指数模型进行精度验证,发现基于NDVI的统计模型在LAI反演过程中具有最高精度,实测小麦数据验证结果为:R2=0.78,MAE=0.380505,RMSE=0.486441。(3)、根据神经网络的超强容错、非线性映射能力,提出基于BP神经网络的LAI反演方法,即依据健康植被对电磁波的反射特性,以植被冠层多光谱数据为输入、LAI为输出构建神经网络反演模型。文章构建两个4层BP神经网络模型针对是否考虑植被冠层方向性特征的情况进行建模分析,其中神经网络1忽略冠层的几何光学特性,而神经网络2考虑了植被冠层BRDF特性,即将Ross Li核参数作为几何光学特性的表征参数应用于神经网络的训练。利用模拟数据和实测小麦数据对两种网络进行精度验证,忽略植被冠层几何光学特性的神经网络1对实测数据的验证精度为:R2=0.80859,MAE=0.45998,RMSE=0.34936;顾及植被冠层BRDF效应的神经网络2精度为:R2=0.82973,MAE=0.44297,RMSE=0.33886。文章得出结论:(1)、在基于经验统计模型反演LAI时,NDVI在12种植被指数中具有更高的反演精度,更好的鲁棒性,反演结果更可靠;(2)、基于植被冠层反射率光谱建立的神经网络反演模型,以其特有的非线性拟合能力使得其反演结果明显优于NDVI统计模型;(3)、考虑植被冠层BRDF效应的神经网络反演精度优于不考虑该效应的神经网络。故文章提出一种基于植被冠层光谱BRDF模型的神经网络LAI反演方法,该方法在一定程度上可以提高LAI的反演精度。
[Abstract]:Leaf area Index (Leaf Area Index,LAI) refers to the sum of all leaf areas above the surface per unit level, which measures the material circulation and energy flow in the energy of the biosphere and the atmosphere. Vegetation growth situation and an important index to monitor global climate change. Although the traditional direct measurement method can meet the needs of practical application in accuracy, it is very destructive and the investment of human and material resources is too large, so it is still very difficult to obtain large area LAI data. Remote sensing monitoring is a non-contact indirect measurement, which is widely used because of its large area, high efficiency, all-weather, non-destructive measurement and so on. In the process of remote sensing image data acquisition, surface reflectivity is often widely used as an important parameter to retrieve the physical and biological characteristics of the ground. However, high precision surface reflectivity products are usually closely related to sensor imaging geometric attitude and atmospheric environment. In this paper, the influence of solar zenith angle and sensor imaging attitude on vegetation spectrum is studied and discussed, and a LAI inversion method based on vegetation canopy BRDF effect is proposed, which uses PROSAIL model to obtain simulation data. This model is input into BP neural network (Back Propagation Artificial Neural Network,BP-ANN as sample set for training to construct inversion model. The main work and corresponding conclusions of this paper are as follows: (1) according to the measured data, reference materials and prior knowledge, the multi-angle parameter database of vegetation radiation transfer model is constructed, and the spectral data of vegetation canopy are simulated on the basis of parameter database. According to the angle information of simulation data, the kernel parameters of semi-empirical Ross-Li (Ross thick-Li transit) model are calculated. (2) 12 vegetation indices of RVI,DVI,NDVI,MSR,EVI1,EVI2,RDVI,ARVI,SAVI,OSAVI,MSAVI,NLI are designed. The empirical statistical model was established with LAI. The accuracy of each exponential model is verified by simulation samples and measured wheat data. It is found that the statistical model based on NDVI has the highest accuracy in the process of LAI inversion. The verification results of measured wheat data are as follows: R2 鈮，

本文编号：2491222