ICESat-GLAS波形与HJ-1A高光谱影像联合反演森林地上生物量的研究

发布时间：2018-03-12 10:39

本文选题：ICESat-GLAS波形　切入点：HJ-1A/HSI高光谱影像　出处：《东北林业大学》2016年博士论文　论文类型：学位论文

【摘要】：传统森林参数获取主要依赖于野外调查,需要耗费大量人力、物力和时间,而且只能获取到点上的数据,不能够及时获取大区域森林参数的空间分布情况。遥感技术的出现和发展弥补了传统估测方法的不足,可以实现对大区域森林参数的快速、无损估测,提高了森林参数的估测精度,能够满足森林资源调查的需要。本研究以吉林省汪清林区为研究区,从ICESat-GLAS(the geoscience laser altimeter system onboard the ice,cloud,and land elevation satellite)波形数据中提取波形参数(波形长度W、地形坡度参数TS、比值能量参数I和差值能量参数ec),利用这些参数实现GLAS波形数据对各森林类型(阔叶林、针叶林和混交林)的森林最大树高、森林郁闭度和森林地上生物量的估测。但由于GLAS光斑呈离散条带状分布,不具有成像性,不能实现森林参数的区域估测,所以本研究引入HJ-1A/HSI高光谱影像,协助GLAS波形数据对森林参数进行空间反演,应用支持向量回归机算法建立GLAS估测的森林参数与HJ-1A/HSI高光谱影像光谱信息间的关系模型,得到森林参数的空间分布图。主要研究结果如下：(1)基于分段主成分分析和波段指数综合法对HJ-1A/HSI高光谱数据降维,选出最佳波段为波段2、波段52和波段107,波段之间的相关性小,信息量大。但由于HJ-1A/HSI数据的空间分辨率低,分类结果不理想(分类精度为70.5%)。将其与CCD数据融合后空间分辨率得到提高,基于支持向量分类机算法,利用融合后影像的波段信息及波段纹理特征信息将分类精度提高到了85.3%。(2)基于GLAS波形提取的地形坡度参数TS,与地形坡度间存在很强的相关性。基于W和TS建立的森林最大树高模型要优于只利用W建立的森林最大树高模型和利用地形指数修正的森林最大树高模型。基于支持向量回归机算法,建立了GLAS估测森林最大树高与HJ-1A/HSI高光谱影像前3个MNF分量的SVR模型,生成森林最大树高分布图,其估测误差的最大值为7.1m,最小值为0.07m,25%的估测误差小于等于0.75m,50%的估测误差在0.75m-2.31m之间,精度较高。(3)对于GLAS波形来说,利用比值能量参数I75所建模型的估测效果优于利用差值能量参数ec所建模型,将I,5和ec联合建立的多变量模型的精度最高。利用HJ-1A/HSI高光谱数据进行森林郁闭度建模时,多变量森林郁闭度模型的估测效果优于单变量森林郁闭度估测效果。HJ-1A/HSI所建多变量森林郁闭度模型的估测效果稍高于GLAS所建多变量模型,因此本研究最终采用HJ-1A/HSI高光谱影像生成森林郁闭度分布图,并对所得森林郁闭度分布图的估测误差进行分析,其中25%的估测误差小于等于0.031,50%的估测误差在0.031-0.126之间,估测精度较高。(4)利用GLAS估测的森林最大树高建立的森林地上生物量单变量模型的估测效果不是很理想,但是将GLAS估测的森林最大树高与HSI估测的森林郁闭度相结合,利用多元线性回归和支持向量回归分别建立各森林类型的森林地上生物量模型,模型的估测效果有了很大的提高,支持向量回归模型的估测效果要优于多元线性回归模型。
[Abstract]:The traditional forest parameter acquisition mainly depends on the field investigation, requires a lot of manpower, material resources and time, and can only access to the data on the spatial distribution can not timely access to large areas of forest parameters. Remote sensing technology and the development of the shortcoming of traditional estimation methods, can be achieved on the large area of forest parameters quickly to improve the accuracy, nondestructive estimation, estimation of forest parameters, to meet the need of forest resource survey. In this study, Wangqing forest area of Jilin Province as the study area, from ICESat-GLAS (the geoscience laser altimeter system onboard the ice, cloud and, land elevation satellite) extracted from the waveform parameters in the waveform data (wave length W, terrain slope parameter TS, the ratio of energy parameter I and the difference between the energy parameters of EC, GLAS) waveform data on all forest types using these parameters (broad-leaved forest, coniferous forest and mixed forest The maximum tree height) forest forest biomass estimation, forest canopy density and forest floor. But because the GLAS spot was discrete strip distribution, not with imaging, can not achieve the estimation of forest parameters, so the introduction of HJ-1A/HSI hyperspectral image, assist GLAS waveform data inversion of forest parameters. Application of support vector machine algorithm to establish the relationship model of forest parameters and HJ-1A/HSI hyperspectral image spectral information GLAS estimation of the spatial distribution of forest parameters. The main results are as follows: (1) the comprehensive method of segmented principal component analysis and band index for dimensionality reduction of hyperspectral data based on HJ-1A/HSI, select the best band to band 2 that band 52 and band 107, band correlation between small and large amount of information. But due to the low spatial resolution of HJ-1A/HSI data, the classification result is not ideal (classification accuracy was 70.5%) with the CCD. After data fusion can improve the spatial resolution, support vector machine classification algorithm based on the information fusion band and band texture feature information image classification accuracy is improved to 85.3%. (2) terrain slope parameter TS GLAS waveform based on the extraction, and terrain slope between the strong correlation between W and TS. The largest tree forest the establishment of the model is better than that of the high model and the high model of maximum tree forest terrain Index Revised by the biggest tree forest based on W. Support vector regression algorithm based on GLAS is established to estimate forest maximum tree SVR model 3 MNF component and high HJ-1A /HSI hyperspectral image, maximum spanning forest tree height distribution map, the maximum estimation error is 7.1m, the minimum value is 0.07m, the estimation error is less than or equal to 25% 0.75m, 50% of the estimated error between 0.75m-2.31m, the accuracy is higher. (3) for the GLAS wave, Li With the ratio of energy parameter I75 model estimation is better than using difference energy parameters of EC model, I, multivariate model 5 and EC jointly established the highest accuracy. The forest canopy density modeling using HJ-1A/HSI hyperspectral data, estimation results of multivariate forest canopy density estimation model is better than the single variable forest the effect of canopy density estimation.HJ-1A/HSI multi variable model of forest canopy density is slightly higher than that of GLAS built a multivariate model, this study adopts HJ-1A/HSI hyperspectral image generation forest canopy density distribution map, and on the estimation of forest canopy density distribution of the error were analyzed, of which 25% of the estimated error of the estimation error is less than or equal to 0.031,50% between 0.031-0.126, estimation accuracy is high. (4) to estimate the effect of biomass of single variable model is established by using high maximum tree forest GLAS estimation of forest land is not Very good, but the maximum tree forest GLAS estimation and HSI estimation combined with high canopy density, biomass models for different forest types of forest land were established by multiple linear regression and support vector regression model to estimate the effect has been greatly improved and the support effect estimation of regression model to better than multivariate the linear regression model.

【学位授予单位】：东北林业大学
【学位级别】：博士
【学位授予年份】：2016
【分类号】：S718.5;S771.8

【相似文献】