【作者】 马孟莉；

【导师】 曹卫星；

【作者基本信息】 南京农业大学 ， 地图学与地理信息系统， 2011， 硕士

【摘要】 作物长势实时无损监测及生产力准确预测,有利于提高作物栽培中因苗管理的精确性。遥感的空间性、实时性可提高作物生长监测与预测能力,本文基于HJ卫星数据,以江苏省如皋市为研究区,研究了基于混合像元分解的水稻信息提取及定量遥感反演方法。具体内容与结果如下：为了解决中低分辨率遥感影像混合像元问题以提高水稻种植信息的提取精度,本文提出了两种适合于多光谱遥感的混合像元分解方法：一是基于图像分块的混合像元分解方法；二是分层多端元混合像元分解方法。第一种方法结合影像空间特征与光谱信息提取端元,即将遥感图像先分块,然后在地物相对简单的各个“小图像”上提取端元,从而弥补端元漏选及错选的缺陷,并实现了增加端元数目进而提高了解混精度；第二种方法是基于层次分类与多端元混合像元分解相结合的水稻面积信息提取方法(Stratified multiple endmember spectral mixture analysis, SMESMA)。层次分类有效降低了地物复杂度,而多端元混合像元分解通过对每一类地物选取多个端元光谱参与解混,克服了“同物异谱”造成的光谱变异问题,两者结合可较大程度提高分类精度。结果显示,第一种方法分类精度为83.65%,kappa系数为0.82；第二种方法(SMESMA)分类精度达到85.78%,kappa系数为0.85,较第一种分类方法略高,将提取的种植面积与如皋市统计局数据相比较,精度超过85%。将两种方法与常规的像元级分类方法(MLC)的精度进行了比较,结果显示本文提出的两种亚像元分类方法精度更高,表明本文提出的方法是适合基于中低分辨率多光谱遥感进行作物分类和面积提取的有效方法。基于水稻信息提取后,利用“纯净”水稻光谱参数与实测水稻叶片氮含量及氮积累建立相关关系,通过统计筛选出与水稻叶片氮素相关性最高的植被指数。结果表明,差值植被指数DVI(4,3)和比值植被指数RVI(4,2)分别与叶片氮含量和氮积累量关系最显著,决定系数分别达到0.74和0.80。此外,基于2010年多时相HJ卫星数据,利用Savitzky-Golay滤波方法有效去除云、异常值及数据缺失的影响,得到较高质量的时序HJ-NDVI和HJ-EVI数据。在此基础上分生育期分别建立了水稻LAI与NDVI和EVI的关系,同时构建了水稻单产与LAI的关系,结果显示,开花时期为水稻最佳估产时相,HJ-NDVI与LAI及LAI与产量关系最好,从而构建了基于“EVI-LAI-水稻单产”的水稻估产模型。通过不同年份试验检验,表明本文所构建的叶片氮素营养监测模型和水稻单产估算模型精度较高,可用于大面积水稻生长状况及产量的监测预测。更多还原

【Abstract】 Real-time and non-destructive monitoring of crop growth and accurate grain yield and qualities predicting could improve traditional crop cultivation and real-time forecasting techniques, and it’s also very important for developments of food security and sustainable agriculture. The ability of monitoring and predicting crop growth has been improved significantly with the spatial and real-time remote sensing images. In this study, the main research are methods of extracting area information based on pixel un-mixing and of quantitative remote sensing reversed based on HJ remote sensing images in Rugao county, Jiangsu. Following are the detail results:To resolve the serious pixel mixing problem of coarse spatial resolution sensors, and improve the cultivation area extraction accuracy of paddy rice, two pixel un-mixing methods were proposed, which were fit for multispectral remote sensed images:Pixel un-mixing based on Image blocking (IBPU) and the stratified multiple endmember spectral mixture analysis (SMESMA). For the first one both spatial feature and spectral information was combined, and the endmember was selected from the "blocked images" containing more simple landscape, which could avoid mistakes of endmember selection caused by deficiency of spectral bands and spectral information of multispectral images, and improve the accuracy of pixel un-mixing. For the later one, the complexity of landscape will be mitigated using stratified classification method, and MESMA represent an alternative approach, in which the number and types of endmembers vary in a per-pixel basis, and overcome the spectral variations within classes. The accuracy of classification was improved significantly by combining these two methods. The result showed that SMESMA had the best classification accuracy of 85.78% and kappa coefficient of 0.85, than the IBPU of 83.65% and 0.82, and both of the two method based on pixel outperformed maximum likelihood classification (MLC). Over all, SMESMA is a mechanism method with higher accuracy, which can solve the "same object with different spectra" phenomenon, IBPU can not solve the it, but has advantages such as simple and high operation efficiency. The results of our study indicated that the two proposed methods were useful and fit for paddy cultivation area extracting with coarse spatial resolution images.Based on the extracting information of paddy rice, the relationships between measured leaf nitrogen content (LNC), leaf nitrogen accumulation (LNA) and "pure" spectral indices of rice were analyzed. For LNC models, DVI(4,2)(differential vegetation index) was the best index with determination coefficient (R2) of 0.74. And for the LNA, RVI(4,2)(ratio vegetation index) was the best index with R2 of 0.80. In addition, the Savitzky-Golay filter was used to remove noise or unusual values such as cloud vapor et al in the HJ images. and then the relationships between time-series HJ-NDVI(normalized difference vegetation index), HJ-EVI(enhanced vegetation index) data and leaf area index (LAI) were analyzed, meanwhile, the relationship between rice yield and LAI was analyzed too, finally, the optimum phase was selected and the rice yield prediction model was constructed base on "EVI-LAI-rice yield".The inspection result with different years showed that this construction of leaf nitrogen nutrient monitoring model and rice yield prediction model is feasible and reliable, which are fit for widespread application.更多还原