【作者】 李晓松；

【导师】 李增元；

【作者基本信息】 中国林业科学研究院 ， 森林经理， 2008， 博士

【摘要】 准确掌握干旱地区的稀疏植被覆盖度是评价干旱地区荒漠化危害程度的基础工作,可为科学防沙治沙决策的制定提供有效的技术支持。植被覆盖度获取的传统方法为地面测量法,该方法仅适用于小尺度植被覆盖度调查,费时、费力,局限性较大。遥感技术的发展为植被覆盖度获取提供了新的技术手段,尤其是为大范围地区植被覆盖度的快速、准确获取提供了可能。然而,受干旱地区特殊性的影响,利用遥感技术精确反演干旱地区植被覆盖度仍面临着以下几大困难:(1)最为主要的是干旱地区植被覆盖稀疏,像元内的植被对像元平均反射率的贡献较小,而且由于干旱地区的土壤有机质含量较低,干旱地区土壤通常颜色较亮,且矿物含量空间异质性较强,裸露的、变化的土壤表面对像元光谱变化的贡献较大,这些因素使得植被在像元中的光谱贡献易于被掩盖;(2)干旱地区开放的植被冠层及明亮的土壤导致明显的多重散射与非线性混合;(3)干旱地区植被的光谱特征与湿润条件下的绿色植被不尽相同,较为突出的是红边特征不明显。因此,如何提高遥感探测干旱地区稀疏植被覆盖度的能力一直是干旱地区植被遥感所面临的最严峻挑战。高光谱遥感是20世纪最后两个十年中人类在对地观测方面所取得的重大技术突破之一,是当前遥感技术的前沿,其利用很多狭窄的电磁波波段(一般波段宽度<10nm)从感兴趣物体获取有关数据,能产生一条完整而连续的光谱曲线,具有常用多光谱遥感无法比拟的优势。然而,受高光谱数据获取能力的桎梏,大尺度植被覆盖度高光谱定量反演方法并没有得到深入的研究及推广,对干旱地区稀疏植被覆盖度的相关研究更是少之又少。星载高光谱数据的民用化从源头上解决了数据获取的问题,因而系统、深入地研究干旱地区稀疏植被覆盖度高光谱定量反演方法,充分挖掘高光谱遥感的潜力,成功扩展高光谱遥感应用领域的工作迫在眉睫。鉴此,本研究以星载高光谱影像Hyperion为数据源,选取典型干旱区甘肃省民勤绿洲—荒漠过渡带为研究区,针对干旱地区稀疏植被覆盖度(< 40%)的定量反演,深入地探讨了植被指数、回归模型法及混合像元分解法反演稀疏植被覆盖度的能力,全面系统地研究了基于星载高光谱数据的稀疏植被覆盖度反演方法,并对不同方法的结果进行了比较分析。研究结果表明:(1)高光谱植被指数探测稀疏植被能力明显优于相应宽波段植被指数,基于Hyperion数据特定窄波段(833.3/640.5nm)的抗大气植被指数(ARVI)探测稀疏植被覆盖表现最佳,线性回归模型的判定系数R2可达0.7294,交叉验证RMSE为5.5488;(2)偏最小二乘回归法、人工神经网络法反演稀疏植被覆盖度的潜力较大。偏最小二乘回归模型中以基于176个波段原始反射率的偏最小二乘回归模型表现最佳,验证样本验证RMSE为3.8197,约为验证样本平均覆盖度的19%。人工神经网络模型中,以精简、不相关却又包含原始影像绝大部分有用信息的主成分为输入的神经网络模型表现最好,验证样本验证RMSE为3.2806,约为验证样本平均覆盖度的16%;(3)基于荒漠植被、假戈壁及流沙三端元的全受限混合像元分解得到的荒漠植被分量与样地实测植被覆盖度吻合程度较高,两者高度相关(R~2为0.9141),29个样地的偏离均不超过5%,RMSE仅为3.0681,约为所有样地植被覆盖度均值的22%;(4)混合像元分解法不需要野外样地数据,可推广性最强,精度相对来说也最高,因而该方法是干旱地区稀疏植被覆盖度反演的最佳方法,表现优于植被指数法及回归模型法。植被指数法精度相对最低,但是其形式简单,推广起来只需要回归实测植被覆盖度数据与最佳植被指数即可。回归模型法精度介于两者之间,可推广性相对来说最差,需要大量样地数据支持。总体来讲,本研究实现了基于星载高光谱影像Hyperion的干旱地区稀疏植被覆盖度的高精度定量反演,充分挖掘了高光谱遥感的潜力,成功扩展了高光谱遥感的又一应用领域。更多还原

【Abstract】 Accurately acquiring sparse vegetation cover in arid regions, which is the foundation work for desertification evaluation, can provide effective technical support for the establishment of scientific anti-desertification decision-making. Field measurement is the traditional way to get the vegetation cover. However, it is time-consuming, laborious, and feasible only to small-scale vegetation cover survey. Remote sensing provides new means for vegetation acquirement, especially for fast, accurate access to vegetation cover in large areas. Now, remote sensing has become a primary means to large-scale vegetation cover investigation. However, owing to arid regions’particularity, retrieving sparse vegetation cover from remote sensing presents some siginificant challenges. The first and most obvious is the fact that because vegetation cover is low, the contribution of vegetation to the area-averaged reflectance of a pixel is small. Furthermore, because of their low organic matter content, soils in arid regions tend to be bright and mineralogically heterogeneous. All of these factors tend to swamp out the spectral contribution of vegetation in individual pixels; Secondly, open canopies and bright soils in arid regions can contribute to significant multiple scattering and nonlinear mixing; Finally, vegetation is spectrally dissimilar to its humid counterparts lacking, most notably, a strong red edge in arid regions. Therefore, improving detection ability of sparse vegetation cover in arid regions, based on remote sensing, is the most serious challenge.Hyperspectral remote sensing, one of the major technological breakthroughs in earth-observing field in the last two decades in the 20th century, is currently the forefront of remote sensing technology, which utilize a lot of very narrow (<10 nm) and continuous electromagnetic wave bands to obtain relevant data information, can produce a complete and continuous spectrum, and with incomparable advantages over commonly used multi-spectral remote sensing. However, acquisition of hyperspectral data is very difficult, therefore, large-scale vegetation cover quantitative inversion based on hyperspectral remote sensing did not receive in-depth study and widely application, related research on sparse vegetation in arid regions is scanty. Recently, satellite hyperspectral data is open to civilian, which solve the hyperspectral data access problem successfully. Therefore, it is high time to systematically, in-depth study sparse vegetation cove quantitative retrieval methods in arid regions based on hyperspectral remote sensing, fully tap the potential of hyperspectral remote sensing, and successfully expand hyperspectral remote sensing application field.Aim to this, we conduct a comprehensive and systematic study to retrieve sparse vegetation cover in Minqin oasis-desert transitional zone based on Hyperion image. Methods we used include vegetation index, regression model and spectral mixture analysis. We exploit each method’s potential to retrieve sparse vegetation cover, and then conduct a comparative analysis among different methods. The results show that:(1) Hyperspectral vegetation indices are significantly better than the corresponding wide-band vegetation indices for sparse vegetation detection, atmospherically resistant vegetation index (ARVI) based on specific Hyperion narrow-band (833.3/640.5 nm) performs best, with a high R2 (up to 0.7294), and a low cross validation RMSE (5.5488);(2) Partial least squares (PLS) regression and artificial neural net (ANN) have great potential to estimate sparse vegetation cover. From validation results with five independent samples, we can find: PLS regression models based on the original 176-band reflectance performs best in all PLS regression models, with a low validation RMSE (3.8197, 19% of mean); ANN models, taking principal components as input, which is compact, not related, but include most original image’s useful information, performs best in all ANN models, with a low validation RMSE (3.2806, 16% of mean);(3) Sparse vegetation fraction, based on fully constrained spectral mixture analysis with sparse vegetation, false Gobi and sand three endmembers, and field measured vegetation cover are highly correlated(R~2 =0.9141). The differences are less than 5% for all samples between them, and RMSE is 3.0681, about 22% of mean;(4) With highest extensibility and accuracy, spectral mixture analysis is the best methods for retrieving sparse vegetation cover in arid regions, and which performs better than vegetation index and regression model obviously.Vegetation index method’s accuracy is relative minimum, but it’s simple, and when applies to other regions, it just need simple, necessary sample investigation. Regression model’s accuracy and maneuverability is both moderate, but its extensibility is worst, and which need a lot of samples to support. Overall, we realize high-precision quantitative retrieval based on Hyperion image of sparse vegetation cover in arid regions, fully tap the potential of hyperspectral remote sensing, and extend another hyperspectral remote sensing successful application field.更多还原