【作者】 彭代亮；

【导师】 黄敬峰；

【作者基本信息】 浙江大学 ， 农业资源利用， 2009， 博士

【摘要】 中国水稻种植面积居世界第二,总产量居世界第一。水稻种植面积的变化、产量丰欠等重要农情信息历来受到各级政府和社会公众的高度重视,是国家制定粮食政策和经济计划的重要依据。及时了解、准确掌握水稻种植面积、产量等对各级政府制定农业生产和农村政策。确保我国乃至世界粮食安全具有重要的意义。国家统计局承担着全国水稻产量的调查职能。经过几十年的努力,形成了较完善的统计调查方法体系。然而,随着我国社会经济的快速发展,政府决策部门、社会公众等统计服务对象对统计的需求提高,以及统计部门面临的问题和压力,如何吸收利用先进技术。逐步改善统计调查手段、降低调查成本、提高调查精度是当前面临和亟待解决的首要问题之一。以遥感为主的3S高新技术具有全覆盖、及时、客观等特点。为改进和提升传统统计手段提供了良好的技术支撑。水稻遥感估产经过多年的发展研究,随着遥感数据的更新以及技术的发展,结合统计局统计抽样调查地块实割实测水稻标准亩产数据,研究统计与遥感相结合的水稻遥感估产,实现统计部门在水稻面积产量监测及预报等业务化运行等方面具有重要意义。湖南省水稻种植面积及总产都居全国首位,且其地形地貌和水稻种植制度在中国水稻种植区具有典型性,因此,本研究以湖南省为研究区,利用MODIS不同数据产品(MOD09A1、MOD09GA、MYD09GA、MOD13Q1、MYD13Q1、MOD15A2、MOD17A2、MOD17A3),在完成水稻遥感估产分区及水稻面积提取的基础上,结合县级、乡镇级统计数据及统计局统计抽样调查地块实割实测数据,以及基于像元水平的MODIS GPP/NPP,研究不同数据完备条件下,水稻遥感估产方法。主要方法与结论如下:(1)基于空间分析与两维图论树算法相结合的湖南省水稻遥感估产分区在分析研究区(湖南省)基本特征的基础上,以水稻种植制度、稻土此、水稻单产水平、地形及县域空间位置作为分区指标,采用空间分析与两维图论树算法相结合的方法,将湖南省分为两个一级区,五个二级区。(2)基于水稻典型发育期光谱特征与多时相MODIS数据湖南省水稻面积遥感信息提取在对MOD09A1数据受云污染的像元进行基于MODIS数据的质量评价(Qualityassessment:QA)信息的相邻时相最大值法插补的基础上,利用水稻移栽期、抽穗期的典型光谱特征、以及相应的EVI与LSWI的关系,排除非水稻植被及其它因素的影响。根据研究区水稻种植制度特点,分别对2000-2008年早稻、晚稻及一季稻面积进行提取。并利用县级统计数据、基于高分辨率遥感影像数字化的水稻田数据、1:1万的土地利用数据,对提取结果进行验证评估。结果显示:湖南省水稻主要分布在洞庭湖流域及湖南省中部地区,相对于以前的研究,本研究的提取精度有较大的提高。(3)基于水稻遥感估产分区、水稻面积提取、县级统计数据的省级水稻总产遥感估算在水稻遥感估产分区、水稻面积提取的基础上,利用1:25万县级行政区划图,提取2000-2008年每年的早稻、晚稻及一季稻对应的各县各时相MOD09A1数据,选取EVI为估产参数。利用水稻生长发育期数据,选择移栽期、孕穗期、抽穗期、乳熟期及成熟期为水稻遥感估产的主要生长期,以未基于分区与基于分区两种思路,建立2000-2007年各县EVI平均值乘以第三章提取的水稻面积的结果(AEVI)与统计总产的一次线性、二次非线性及各生育期逐步回归模型。通过误差分析,选择水稻总产最优遥感拟合模型。在此基础上,利用基于2000-2007年数据得到的水稻总产最优遥感拟合模型预测2008年的省级水稻总产。结果显示:基于分区的估产模型要比未基于分区的水稻遥感估产模型要好,且生育期主要集中于孕穗期到乳熟期,说明这段时间是水稻产量形成的最关键时期。另外。二次非线性模型及逐步回归模型的结果要好。误差分析结果显示,水稻遥感估产模型的拟合结果比预测结果要好,两者省级相对误差都小于5%。(4)基于统计局统计抽样调查地块实割实测数据、MOD13Q1与MYD13Q1EVI的省级水稻单产遥感估算基于国家抽样县抽样地块的实测数据是统计部门计算水稻单产的基础数据,本研究根据地块的空间位置,提取各地块所在地的3×3像元网格对应的MOD13Q1与MYD13Q1相结合的EVI数据,选择移栽期、孕穗期、抽穗期、乳熟期及成熟期的EVI,建立地块实测标准亩产与EVI的一次线性、二次非线性及各生育期逐步回归模型。通过误差分析,选择水稻单产最优遥感拟合模型。在此基础上,利用上一年的水稻总产最优遥感拟合模型预测下一年的省级水稻单产。根据水稻面积提取结果,得到省级水稻总产。与统计值进行比较,结果显示,在省级水平上,基于地块的水稻单产与总产遥感拟合与预测结果相对误差都接近5%。(5)基于乡镇级统计数据、MOD13Q1与MYD13Q1 EVI的的县级水稻总产遥感估算选择醴陵市为研究区,分析基于乡镇级统计数据的水稻遥感估产。选择分别由Terra与Aqua卫星获取得到的MOD13Q1与MYD13Q1 EVI作为水稻估产遥感指标。其空间分辨率约为250 m。时间分辨为8 d。选择移栽期、孕穗期、抽穗期、乳熟期及成熟期的遥感数据,建立EVI乘以水稻面积的结果(AEVI)与乡镇级水稻统计总产的一次线性、二次非线性及各生育期逐步回归模型。通过误差分析,选择水稻总产最优遥感拟合模型。在此基础上,利用上一年的水稻总产最优遥感拟合模型预测下一年的醴陵市水稻总产。估产结果显示,各拟合模型相对误差都小于0.1%。预测结果误差相对于拟合结果较大,但相对误差仍然小于5%。(6)基于像元水平的MODIS GPP/NPP数据的水稻遥感估产、及像元内水稻面积比对估产精度的影响在前人研究的基础上,总结得到基于MODIS GPP/NPP水稻估产方法。利用MODIS 8d的、年的GPP/NPP产品及产品使用说明,估算8 d的NPP,并利用此结果与8d的GPP产品分别进行基于GPP/NPP的水稻遥感估产。借助MODIS GPP/NPP产品生成1×1km网格。研究像元内不同水稻面积比对估产结果的影响。研究结果表明:利用估算年NPP的方法、8 d的LAI和PSNnet来估算每8 d的NPP的结果比MODIS年NPP值较大。基于MODIS NPP的水稻遥感估产结果相对于统计值较小.而基于MODI GPP水稻遥感估产与统计值接近。其估产精度随着像元内水稻面积的增加而减小。像元内水稻面积超过80%,其相对误差小于20%或10%。综上所述,本研究的主要结论如下:基于县级、乡镇级、实测地块数据、MODISGPP数据的水稻遥感估产的结果基本上都能达到95%的精度。从提高精度方面考虑,在乡镇级统计数据具备的前提下,利用基于乡镇统计数据及MOD13Q1、MYD13Q1EVI的方法是最优选择。从水稻估产业务化运行角度来考虑,在地块实测产量及空间位置数据都具备的前提下,利用基于统计局统计抽样调查地块实割实测数据及MOD13Q1、MYD13Q1 EVI的方法较好。在统计局统计抽样调查地块实割实测数据不具备的情况下,利用分区与水稻面积提取结果、以及县级统计数据的水稻遥感估产方法。而对于大面积水稻种植区,利用基于像元水平的MODIS GPP/NPP估产方法较好。更多还原

【Abstract】 China is a large agricultural country,rice area and total output occupies second and first place in the world,respectively.All levels Chinese government and social public pay high attention to the variation of rice area and yield,which information is the very important for food economic policy decision.Therefore,it is very essential to know about the rice area and yield in time.Chinese Bureau of Statistics undertakes the nationwide rice yield investigation.After several decades’effort,a statistic and investigation system has been formed.However,with the development of social economy,government decision department,social public and some other statistic service objects have increased their demands for the statistic results,as well as other problems and pressures,it is urgent to improve the traditional investigating method using the advanced technologies.3S technologies,especially remote sensing,features the large area coverage,in-time, objectivity and so on,provide very good tools for improving the traditional statistic method.Remote sensing rice yield estimation has been studied for several years. However,with the new satellite data and technologies updated,the study of rice yield estimation approach using the integrated temporal and spatial resolution remote sensing data is very helpful to implement the rice area and yield estimation on vocational operation.Hunan province was selected as the study area in this thesis,where the rice area and total yield occupy the first place in China,and features the typical topography and rice cropping system.After rice yield estimation division and rice detection,the purpose of this study is to build different models and estimate rice yield using different MODIS products(MOD09A1,MOD09GA,MYD09GA,MOD13Q1,MYD13Q1,MOD15A2, MOD17A2,MOD17A3) at the different level.The summary of the major chapters in this thesis as follows:1 The study of rice yield estimation division based on spatial analysis and two-dimensional optimal tree clusters in Hunan provinceRice planting system,unit yield of rice,the ratio between rice and total land area,and the ratio between plain and total land area were selected as the main factors for rice yield estimation division.Two first-grade zoning,and five-grade zoning in Hunan province were got based on the integration of spatial analysis and two-dimensional optimal tree clusters,2 Detection and estimation of paddy rice area based on rice typical phenology spectral characteristics and multi temporal MODIS data in Hunan provinceThe periods of rice transplanting and heading were selected as the key stages,the pixel of MOD09A1with clould contamination was filled by QA information and adjacent maximum fitting method.Using the relationship between EVI and LSWI during the transplanting and heading periods,non-paddy rice plant and other disturbing factors were removed,the early,late and singlerice from 2000 to 2008 were extracted.Our results were validated with finer resolution(2.5m) Satellite Pour l’Observation de la Terre 5 High Resolution Geometric(SPOT 5 HRG) data,land-use data at the scale of 1/10000 and county-level statistical rice area.The results showed that three paddy rice crop patterns could be discriminated and their spatial distribution quantified.We show the area of single crop rice have increased annually and almost doubling in extent from 2000 to 2008,but unique declines in the extent of early and late paddy rice.These results were more accurate than previous satellite-based methods.3 Provincial rice total yield estimation based on the results of division and rice detection using remote sensing and county level statistic dataBased on rice yield estimation division and rice area detection,MOD09A1 EVI corresponding the rice area detection from 2000 to 2008 was extracted using the administrative map at the scale of 1/250 000 at the county level.Transplanting,booting, heading,milky and harvest stage were selected as the key periods for rice yield estimation,EVI multiplied by county rice area and county rice total yield were took as the independent and dependent variables respectively,and then different models were made based on division results and not using MODIS and statistics data from 2000 to 2007.The optimal rice yield estimation models were selected after errors analysis,and then the rice forecasting yield in 2008 were obtained.The results showed that the accuracy of rice yield estimation based on division was better than without division,and almost all optimal models focused on the period from booting to milky,and the quadratic nonlinear and stepwise models are better than linear models.The estimation yield had a significant positive correlation with statistical data,and the relative errors for rice estimation and forecasting models were less than 5%.4 Provincial rice unit yield estimation using MOD13Q1,MYD13Q1 EVI and plot level measured dataOne of the very important sources for statistic data for Chinese Statistics Bureau is the measured data from sampling plot.3×3 window size MOD13Q1 and MYD13Q1 combined EVI corresponding with those plots in Hunan province were extracted. Transplanting,booting,heading,milky and harvest stage were selected as the key periods for rice yield estimation,and the different type models were built.The optimal rice yield estimation models were selected after errors analysis,and then the rice forecasting yield were obtained.According to the results of rice detection,the provincial rice total yield was obtained.Results showed that the relative errors of forecasting rice yield and were less than 5%.5 County rice yield estimation using MOD13Q1,MYD13Q1 EVI and township level statistic dataAccording to remote sensing rice yield estimation based on county level statistic data, one mountain county(Liling) with a large error was selected to further analysis the rice yield estimation using township level statistic data.MOD13Q1 and MYD13Q1 combined EVI corresponding with the rice area based on 1/10000 scale land-use data was extracted. Transplanting,booting,heading,milky and harvest stage were selected as the key periods for rice yield estimation,EVI multiplied by township rice area and township rice total yield were took as the independent and dependent variables,and then different type models were built.The optimal rice yield estimation models were selected after errors analysis,and then the rice forecasting yield were obtained.The results show that the relative errors for rice estimation and forecasting models is less than 0.1%and5%, respectively.6 Rice yield estimation based on MODIS GPP/NPP data at the pixel level,and the impact of the rice area percentage in one pixel on the accuracy of rice yield estimationBased on previous researches,algorithms for rice yield estimation using MODIS GPP/NPP data were studied.8-day NPP was calculated using 8-day LAI,PSNnet data and MODIS17 user’s guide,and then rice yield estimation was performed using 8-day MODIS GPP/NPP and rice growth period data.We used MODIS GPP image to create 1×1km grid,and calculated the percentage of paddy rice area in each grid,and then the impact of different percentage of paddy rice area in each grid on rice yield estimation result was studied.The result showed that:there is a significant positive correlation between 8-day estimation NPP results and MODIS annual NPP product.(P＜0.01),and the rice yield estimation based on NPP data was low compared to statistic data,but the results of MODIS GPP is better.The relative errors between rice yield from estimation and statistic data decrease with the percentage of rice area in the 1×1km increasing,and they are less than 20%,even 10%when the percentage is larger than 80%According to the above researches,the main conclusions are as follows:the relative error of the rice yield estimation based on county,township level statistic data,plot measured data,and MODIS GPP was less than 5%.However,for the sake of improving rice yield estimation accuracy,the models based on the integration of Terra and Aqua data at the township level are the best choice.Considering rice area and yield estimation on vocational operation,the methods based on the integration of Terra and Aqua data and measured data are the best choice.If the plot measured data aren’t available,the rice yield estimation methods,with the support of zoning and rice area detection,and based on the county level statistic data are perfect.For the large paddy rice area,the rice yield estimation algorithms based on MODIS GPP/NPP at the level of pixel are dependable.更多还原