【作者】 穆悦；

【导师】 冯益明；

【作者基本信息】 中国林业科学研究院 ， 水土保持与荒漠化防治， 2017， 博士

【摘要】 戈壁是干旱环境中的一种独特地貌景观。戈壁表面砾石形貌参数研究对于研究戈壁形成演化过程和风沙活动影响,以及开展戈壁地区生态保护和经济开发都具有重要意义。目前在戈壁区开展的大范围、高精度的砾石形貌特征参数研究较少,究其原因主要是地面测量精度高,但受人工测量效率的限制难以获得大量的样本;遥感估算覆盖面积大但是由于遥感影像空间分辨率较低,难以直接测量砾石形貌特征。随着近地面遥感技术的发展,在戈壁表面获取较大覆盖范围、超高地面分辨率的数字图像越来越便利。因此,利用计算机图像模式识别领域先进的图像分析技术发展针对高空间分辨率图像的图像分析技术具有重要的科学意义和应用前景。然而,迄今为止,基于图像自动计算砾石形貌参数的方法在戈壁的形成演化研究中尚未见到报道。因此,本研究提出一种基于图像,利用机器学习算法快速、准确测量戈壁表面砾石覆盖度、粒径、形状比率和方位角等形貌参数的新方法,并利用地面实测和人工矢量化地面摄影数字图像验证了新方法的精度。应用该方法以中国新疆维吾尔自治区哈密市天山南麓洪积扇(93°0′0″E,43°10′0″N)为研究区域,利用野外地面摄影、近地面无人机遥感和卫星影像三种不同尺度数据源,实现了天空地一体化的戈壁表面砾石形貌特征和空间分布规律研究。主要研究成果如下:(1)提出了一套基于决策树图像分类算法、分水岭图像分割算法,快速、准确测定戈壁表面砾石覆盖度、粒径、形状比率和方位角等形貌特征的新技术(Morphological characteristics gain effectively technique,McGET)。通过与地面测量和人工矢量化测量的形貌特征参数进行比较,评估了McGET方法的精度。结果显示:与地面测量结果比较,McGET计算的大砾石(粒径不小于对应样方地面测量粒径最小值)粒径均值与地面测量结果非常接近(y=1.00x+2.74,R2=0.89,P<0.001)。与手动勾绘结果比较,McGET计算的砾石覆盖度精度范围在92.6%~99.9%;在戈壁表面不存在大量叠置的砾石的情况下,McGET计算的粒径均值精度大于69%;McGET计算的砾石形状比率均值精度在88%~99%;在戈壁表面不存在大量近圆形砾石的情况下,McGET获得的砾石方位角分布与手动勾绘一致。更重要的是,McGET大大缩短了野外地面测量及室内图像处理花费的时间。因此,McGET可以同时处理大量戈壁样方图片,快速、准确获取多类型的砾石形貌参数数据。(2)McGET技术在样方(1m×1m)、景观(3 ha)和区域(300 km2)尺度上的适用范围有所差异。在样方尺度上,McGET可以提取到单个砾石的所有形貌特征,包含砾石覆盖度、粒径、形状、方位角等,只不过受图像地面分辨率和覆盖范围的限制,样方尺度上提取的砾石粒级范围多在4mm到256mm左右,比较适合粒级较小的戈壁表面砾石研究。在样区尺度上,McGET也可以提取到单个砾石,只不过受图像地面分辨率限制,提取的砾石粒径在32mm以上,比较适合粒级较大的戈壁表面砾石研究。在区域尺度上,McGET可以获取像元内砾石特征的均值,适合通过遥感反演进行大尺度上砾石粒径的分级研究。(3)应用McGET技术结合运动恢复结构和遥感反演技术,分别以地面数字图像、无人机高分辨率图像和卫星影像为数据源,在样方、景观和区域尺度上上对戈壁表面的砾石形貌特征进行了估算。样方尺度上,砾石覆盖度均值约为75%;砾石呈现单峰左偏态分布,砾石粒径均值约为15mm;砾石形状比率均值约为1.57。景观尺度上,自扇心到靠近扇缘,三个样区的砾石覆盖度分别为34.22%,26.85%,21.88%;砾石粒径均值分别为130,95,78 mm。区域尺度上,以Landsat 8 OLI地表反射率产品为数据源,结合地面样方调查,建立了戈壁表面砾石粒径遥感估算模型,获得了较好精度(R2=0.549)。(4)本研究通过综合样方、景观及区域尺度上砾石特征的测定结果,分析了砾石特征随海拔、坡度等因素的变化,得到了景观和区域尺度上戈壁表面砾石的空间分异规律。在景观尺度上,沿海拔变化方向,砾石覆盖度和砾石粒径变化与海拔关系不大,可能受到植被分布和局部地形的影响。以地貌单元为单位分析,在靠近扇心区域,平地砾石覆盖度和粒径都明显高于坡地,扇中区域二者差异不大,靠近扇缘区域平地上砾石覆盖度和粒径都略低于坡地上。自扇心到扇缘,平地上砾石覆盖度和粒径的下降速度由快变缓,坡地上砾石覆盖度和粒径的下降速度相对平地较为均匀。在进行景观尺度的戈壁表面砾石特征空间分布研究时,局部微地形的影响不可忽略。在区域尺度上,从洪积扇扇心到扇缘,戈壁表面的砾石覆盖度和砾石粒径都逐渐减小,而砾石形状比率略微增大。线性回归分析显示,海拔每升高100m,砾石覆盖度显著增加了2.2%,砾石粒径均值显著增加了0.5mm左右,而砾石形状比率下降了0.004左右。沿着海拔变化方向,砾石粒径均值逐渐减小且在扇心和扇缘区域下降速度较快,随水平距离增加分别为1.83mm/km和0.31mm/km,在扇中区域下降速度较慢,为0.15mm/km。同一海拔梯度上,在接近洪积扇中轴线的位置砾石粒径较高,而两侧粒径较小;其中扇心区域,二者差别最大,大于3mm,扇中区域二者差别最小,接近2mm。在区域尺度上,砾石粒径变化受宏观地形下的坡度影响较大。更多还原

【Abstract】 Gobi is a unique landscape in arid environment.Study on morphological characteristics of gravels on Gobi surface has important significance on research of formation and evolution of Gobi and the influence of sand transportation,as well as the regional ecological protection and economic development.Currently,the study of characteristic parameters of gravels covering large area with high-precision is few.The main reason is:(1)the accuracy of ground measurement is high,but is difficult to get many samples as the efficiency limitation of manual measurement,(2)remote sensing estimation covers large area,but is difficult to acquire frequency distribution of gravel feature.With the development of the close-range photogrammetry and remote sensing technologies,it becomes more and more convenient for acquirement of digital images with wide coverage and high ground resolution.Therefore,it is of great scientific significance and application prospect to develop high-resolution image analysis method by combining with the advanced image analysis and processing technology in the field of computer information.So far,automatic method for gravel morphological characteristics calculation based on machine learning in the formation of the Gobi evolution studies haven’t been seen in the report.Therefore,this study put forward an image-based,rapid and accurate measurement method for gravel morphological characteristics calculation,e.g.gravel coverage,diameter,aspect ratio and orientation,using machine learning algorithm.The accuracy of this method was verified by field measurement and manual digitization.Applicating this method,using three kind of data sources,i.e.field photographs,images near the ground collected by Unmanned Aerial Vehicle(UAV)and satellite images,has completed the research of morphology characteristics and spatial distribution of gravels on Gobi surface from air-to-ground,with pluvial fan(93° 0’ 0 " E,43° 10’ 0" N)at south slope of Tianshan mountain in Hami,Xinjiang as study area.The main results are as follows.(1)A rapid,accurate morphological characteristics determination method(Morphological characteristics gain effectively technique,Mc GET)of gravels on Gobi surface was developed by combining the decision tree algorithm and watershed transformation.The accuracy of morphological characteristics obtained by McGET was evaluated by compared with field measurement and manual digitization.Results of comparative analysis between McGET and field measurement show the mean gravel diameter measured by field measurement agreed well with that calculated by McGET for large gravels(y = 1.00x+ 2.74,R2 =0.89,P<0.001).Results of comparative analysis between McGET and manual digitization show the gravel coverage calculated by McGET is of very high accuracy,ranged from 92.6% to 99.9%.Unless the surface is covered by a large number of overlapped gravels,McGET can obtain accurate mean gravel size data with accuracy not lower than 69%.Mean gravel aspect ratio obtained by McGET is of high precision ranging from 88% to 99%.Unless the image contains a large number of round gravels,McGET can obtain consistent gravel orientation distribution with manual digitization.More importantly,McGET significantly shorten time cost on obtaining gravels morphological characteristics in field and laboratory.Therefore,McGET can handle a large number of samples in a short time,obtaining the accurate and diverse gravel morphological characteristics.(2)On the scales of quadrat(1m × 1m),landscape(3 ha)and regional(300 km2),the range of application of McGET are different.On the quadrat scale,McGET can get all the morphological characteristics of a single gravel,including gravel coverage,size,shape,orientation,etc.Just limited by the scope of image cover,gravels extracted on quadrat scale is around 4 mm to 256 mm,seeming more suitable for the study of Gobi surface covering fine gravels.On landscape scale,single gravel also can be extracted.Restricted by ground resolution,gravel size is above 32 mm,seeming more suitable for coarse gravel extraction.On the regional scale,only the mean value of gravel characteristics within the pixel can be calculated,which is suitable for large scale gravel size classification research through remote sensing inversion.(3)With the usage of McGET,Structure from motion(SfM)and remote sensing technology,taking images captured on the ground,high resolution images captured by UAV and remote sensing image as data source,morphology characteristics of gravel on Gobi surface were estimated on quadrat scale,landscape scale and regional scale,respectively.On the quadrat scale,the gravel coverage was with mean value of 75%;the gravel diameter was with mean value of 15 mm,unimodal and left skewed distribution;the gravel aspect ratio was mean value of 1.57.On landscape scale,from the center to the edge of the fan,the gravel coverage of three sample zones were 34.22%,26.85%,21.88%;the mean gravel diameter of three sample zones were 130,95,78 mm.On regional scale,with gravel characteristics determined from samples and the surface reflectance of 7 OLI bands as the data source,estimation model of the gravel size was built based on remote sensing data with high precision(R2 = 0.549).(4)Through synthesis analysis of gravel character determination results on quadrat scale,landscape scale and regional scale,changes of gravel characteristics with the factors such as altitude,slope,were analyzed.Then spatial variability of gravel characteristics on landscape and regional scale were obtained.On the landscape scale,along the direction of elevation change,gravel coverage spatial distribution had little relationship with elevation,but may be affected by vegetation distribution and local topography.With geomorphic unit as analysis unit,in the region near the center of the fan,gravel coverage and diameter on the flat were significantly larger than on the slope;in the region near the middle of the fan,the difference of gravel coverage and diameter on the flat or slope is not big;in the region near the edge of the fan,gravel coverage and diameter on the flat were both slightly less than on the slope.On spatial variability study of gravel characters on landscape scale,the influence of the local topography can not be ignored.On regional scale,from the center to the edge of the pluvial fan,gravel coverage and diameter on the surface of Gobi both decreased,while the aspect ratio increased slightly.Linear regression analysis showed that,as the elevation increased 100 m above sea level,gravel coverage increased at the rate of 2.2%,mean gravel diameter increased at the rate of 0.5 mm,and mean aspect ratio slightly decreased at the rate of 0.004.Along the direction of elevation change,gravel and the mean diameter decreased and dropped faster in the region of the center and edge of the fan.With the increase of horizontal distance,the rate of descent was 1.83 mm/km,0.31 mm/km and 0.15 mm/km.On the same elevation gradient,gravel diameter near the central axis of the pluvial fan axis is larger,and near both sides were smaller.In the center of the fan,the difference of gravel diameter between on the central axis or sides of the fan was biggest,which was more than 3 mm.While it was smallest in the middle of the fan,which was near 2 mm.From the center to the edge of the fan,the gravel coverage and diameter on the flat dropped from fast to slowly,while on the slope dropped relative gradually.On the regional scale,the spatial variability of gravel diameter was greatly influenced by slope from macrorelief.更多还原