【作者】 琚存勇；

【导师】 蔡体久；

【作者基本信息】 东北林业大学 ， 生态学， 2009， 博士

【摘要】 土地荒漠化是指由于恶劣的自然因素和不利的人类影响所造成的干旱、并干旱和干燥的半湿润地区的土地退化,是全球广泛关注的重大生态环境问题之一。荒漠化监测与评价是荒漠化研究的核心内容之一,只有及时掌握荒漠化时空变化信息,方能制定并实施有效的宏观管理,做好荒漠化防治工作,保护环境资源可持续发展。然而,到目前为止还没有一种大家公认的正确评价方法。计算机硬件技术与遥感技术的迅速发展,又为这一目标的实现提供了机遇。本文选取在我国颇具代表性的毛乌素地区为研究对象,以多源遥感数据融合与地形地貌分类为基础,通过仿真建立一定数量野外样地调查数据与遥感数据的对应关系,实现土地荒漠化的定量估测。主要研究成果有:1.比较了高空间分辩率SPOT5全色影像与多光谱Landsat5 TM影像不同融合方法的融合质量。在主成分变换融合、比值变换融合和乘积变换融合这三种方法中,乘积变换融合图像即较好地继承了TM影像的光谱质量又最大程度地融入了SPOT全色影像的空间纹理细节。融合影像与TM影像的平均光谱质量差异仅为22.886,而融合影像与SPOT全色影像高通滤波后的相关系数则高达0.949。2.比较了融合影像与原多光谱影像的地貌监督分类效果,发现乘积变换融合图像分类效果最好,165个调查样地总体分类精度达到82.42%,Kappa系数达到0.616。另外,对不同地貌类型调查样地对应的遥感数据进一步分析发现,缨帽变换的湿度指数与绿度指数的比值可以较好地提取道路与水体,而主成分变换的第二成分可以区别沙地与梁地,利用这种分层分类法总体分类精度提高到85.85%,Kappa系数达到0.635。3.针对多变量筛选问题,比较了传统的基于最小二乘穷举法的复相关系数准则与基于偏最小二乘法的Bootstrap方法和变量投影重要性准则。三者所选变量有很大不同,根据bootstrap和变量投影重要性准则所选变量建立的模型精度略有提高,并且筛选变量所用时间急剧减少。变量投影重要性准则虽然是一种定性评价变量重要性的方法,但根据这一准则筛选的变量符合参数节俭原则,并且意义明确。4.研究了地貌类型对植被盖度估测模型的影响,按地貌类型分别建立模型估测植被盖度比不考虑地貌影响时RMSE（均方误差,Root Mean Square Error）减少了3%,而相对RMSE减少了13%。说明地貌分类有助于提高植被盖度的估测精度。5.考虑到变量间共线性的存在,引入了岭回归、偏最小二乘回归和广义回归神经网络模型估测生物量,它们都能克服共线性对模型的不利影响。估测梁地植被生物量时,偏最小二乘回归方法精度最高,RMSE为64.39g.m^-2,相对RMSE为57.68%;估测沙地植被生物量时,广义回归神经网络模型精度最高,RMSE为53.59 g.m^-2,而相对RMSE仅为21.75%。综合来看,沙地比梁地的生物量估测精度高,可能是相对沙地而言,梁地土壤背景复杂造成的。6.分析了土壤水分、植被盖度与生物量与土地荒漠化程度的相关关系,三者中以植被盖度与荒漠化程度的相关性最大,以土壤水分的相关性最小。以植被盖度与生物量为自变量建立荒漠化程度估测模型,正确预报率平均达到81.2%,预报偏差不超过一个等级。考虑到遥感估测植被盖度与生物量有一定误差,根据误差传播定律,必然会带入荒漠化程度估测模型,降低荒漠化估测精度。因此,建立了以遥感因子为自变量的荒漠化程度估测模型,平均正确预报率达到83.9%,预报偏差也不超过一个等级。总之,以图像融合与地貌分类为基础,通过一定数量野外调查样地与遥感数据建立对应关系,可以实现基于遥感数据的土地荒漠化的定量估测,实现荒漠化程度估测的自动化与可视化。更多还原

【Abstract】 Desertification,land degradation in arid,semi-arid and dry sub-humid regions resulting from abominable natural conditions and uncomfortable human activities,is a global ecological and environmental problem.Monitoring and assessment of desertification is a key content of desertification research.Only if spatial and temporal information of desertification changes is obtained in time,effective management can be planed and implemented,and hence prevention and control of desertification can be achieved and natural resources can be protected and developed sustainably.However,there is no consensus on the proper way to assess desertification.But the rapid development of computer hardware technology and remote sensing technology provide an opportunity to achieving this goal.Based on image merging of multiple source and topography classification,the relationships between field inventory data and remote sensing data were conducted and achieved quantitative estimation of land desertification in a typical semi-arid region Mu Us in inner Mongolia.The main research achievements are as following:The quality of merging high resolution SPOT panchromatic image and multispectral TM images using three methods was compared.Among three methods principle component transform,Brovey transform and multiplicative transform,the last one not only inherited more spectral quality of TM images but also merged more spatial context of SPOT panchromatic image just as the average differences between the pixel value of the merged image and corresponding original TM(the registered and the resampled one ) is 22.886 and the correlation coefficients of the high-pass filtered merged image with the high-pass filtered SPOT panchromatic image is 0.949.The image merged by the means of multiplicative transform was the best one among three merged images and original TM image to classify the topography and the overall classification precision was 82.42%and the Kappa coefficient was 0.616.In addition,according to further analyzing remote sensing data corresponding to field inventory data,the ratio of green index to humid index of tasseled cap transform can be used to recognized water and road from other classes and meanwhile the second component can be used to divide highland from sandy land. In terms of the hierarchical classification method the overall classification precision rise to 85.85%and the Kappa coefficient rise to 0.635.In view of selecting variables,multi-correlation coefficient rule based on traditional least squares estimate was compared with the Bootstrap method and the rule of Variable importance in Projection based on partial least squares estimate.Selected variables were different due to used methods and those variables selected by the latters induced to a higher precision of model and took very limited time.Furthormore,those variables selected by VIP rule were interpreted distinctly though the VIP rule is a method applied in quantitative assessment.The effect of Topography types on percent vegetation cover estimating was analyzed.The root mean square error(RMSE)of estimated fractional cover decreased 3%and relative RMSE even decreased 13%due to topography classification.This illuminated that topography classification contributed to improving the estimate precision of percent vegetation cover.Considering that collinearity existed generally among variables the methods ridge regression,partial least square regression(PLSR) and general regression neural network (GRNN) were introduced to estimated biomass due to their robustness of collinearity data.The PLSR among three methods had the best precision of estimating biomass in highland and the RMSE was 64.39 g.m-2,relative RMSE was 57.68%.However The GRNN methods had the best precision of estimating biomass in sandy land and the RMSE was only 53.59 g.m-2, relative RMSE was 21.75%.The precision of estimated biomass in sandy land is better than that in highland.The relationships of soil water content,percent vegetation cover and biomass with land desertification were examined and the correlation of percent vegetation cover with desertification is the biggest,the correlation of soil water content with desertification is the least.The model introducing percent vegetation cover and biomass properly predicted 81.2% of sample plots of land desertificati0n degree and the deviation is lower than one grade.But percent vegetation cover and biomass estimated according to remote sensing data included some error unavoidably,in terms of the error spread rule these errors would be introduced into the model and decrease the predicted precision of land desertification degree.Hence the model only included remote sensing data factors as independent variables was derived to estimate land desertification degree.The result showed the ratio of properly predicted sample plots got an average level of 83.9%and the deviation was less than one grade.In a word,based on merged image and topography classification,quantitative estimate and visualization of land desertification degree can certainly be achieved according to the relationship derived from field inventory and corresponding remote sensing data.更多还原