【作者】 陈思宁；

【导师】 申双和； 赵艳霞；

【作者基本信息】 南京信息工程大学 ， 应用气象学， 2012， 博士

【副题名】以东北地区玉米估产为例

【摘要】 粮食问题是农业生产的重中之重,精确、实时的作物长势监测信息及产量预测信息是农业管理和粮食安全的重要保障。作物生长模拟模型是监测作物生长发育、估算作物产量的有力工具,然而将基于站点研发的作物模型应用于大范围区域作物长势监测和产量预报时面临难以获得区域尺度上的作物参数、农业管理、种植布局信息缺乏等问题,也不易解决模型在区域尺度上的适用性验证问题。随着遥感技术的发展,遥感信息已经成为完善作物模型模拟的主要数据支撑。由于作物生长模型是对实际作物生长发育过程的简化,不仅模型本身存在误差,初始条件、边界条件同样也存在误差。因此,随着作物模型的向前积分,模拟值与真实值之间的差距会越来越大。遥感可以反演作物冠层要素的空间分布,即提供了一种作物冠层要素的客观观测值。因此,作物模型与遥感信息的相互结合十分重要,一方面利用作物生长模型来约束遥感反演模型,另一方面利用遥感数据来调整作物模型的运行轨迹,使积累的误差得到“释放”。最大限度的利用不同来源、不同空间与时间分辨率数据,并将它们有机地融合,更好的表达各种时空尺度上的作物生长过程。数据同化方法的出现和渐趋实用化,为我们达到这一目标提供了一条可行的途径。针对目前国内外研究和应用中存在的问题,并考虑到区域作物估产的需要,本文首先对研究区进行作物分类研究并通过两种玉米制图方案对比综合的方法提取玉米种植区。然后,在引进国外基于集合卡尔曼滤波(EnKF)构建的遥感信息-作物模型耦合模型(PyWOFOST)的基础上,通过对该模型的修改与完善,建立了以LAI为结合点,适用于东北玉米种植区的同化模拟模型,并使用MODIS LAI作为外部同化数据进行同化模拟,将同化模拟结果与常规作物模型(WOFOST)模拟结果比较发现同化后的模拟结果更接近实测值,从而揭示了基于数据同化方法的作物模拟的优势。同时,重点分析了遥感观测(MODIS LAI)和模型参数(TSUM1)的不确定性对同化模拟结果的影响。最后,在玉米制图的基础上,利用完善后的PyWOFOST模型实现了区域尺度上的玉米估产,并利用东北三省各地级市的产量统计数据验证了区域估产的结果,进一步揭示了基于集合卡尔曼滤波同化遥感信息进行作物估产的适用性和可行性。论文的主要研究工作和初步结论如下：1.基于波谱分析方法的作物分类及玉米制图研究本文基于波谱分析方法对研究区进行作物分类制图,并使用两种制图方案对比综合的方法提取玉米种植区。1)考虑到区域作物估产的需要及实测资料的限制,本文使用MODIS陆表产品(包括陆地覆盖类型数据、NDVI数据及反射率数据),通过提取各类作物(大豆、玉米、水稻及小麦)纯净像元的NDVI时间序列曲线,利用波谱分析方法(SAM)进行作物分类研究,在此基础上提取玉米种植区(制图方案1)。2)大豆,玉米及水稻的遥感分类面积和统计面积的相关系数分别为0.858、0.715、0.927,确定系数R2分别为0.770、0.710、0.686,表明基于波谱分析方法获得的作物类型分布图精度比较理想。特别地,对玉米的制图精度而言,黑龙江、吉林、辽宁玉米的分类面积和统计面积的相关系数分别为0.9527、0.6528、0.3462,确定系数分别为0.8291、0.813、0.6883,辽宁省的玉米制图精度不够理想,因此,根据研究区内作物耕种的实际情况,本文采用由中国科学院地理科学与资源研究所(简称中科院地理所,IGSNRR)提供的1km的陆地覆盖类型数据并结合海拔高程数据,提取高程介于0-400m之间且旱地类型占像元面积80%以上的区域作为辽宁省的玉米种植区(制图方案2)。2. PyWOFOST模型在东北玉米估产中的适用性验证通过将PyWOFOST的同化模拟结果与WOFOST的模拟结果及实测结果相比较(主要是玉米LAI、产量、发育期的比较)发现,同化后的模拟结果更接近实测值：1)20个未受灾害影响的农气站玉米产量同化前的模拟误差及在TSUM1的不确定性为0、10、20、30℃时的同化后模拟误差分别为14.04%、12.71%、11.91%、10.44%及10.48%,玉米模拟产量与实测产量的相关系数为0.681,确定系数为0.597,而TSUM1的不确定性为0、10、20、30℃时同化后的模拟产量与实测产量的相关系数均达到0.7以上,确定系数分别为0.631、0.678、0.724及0.697,可见,同化了外部观测数据后的产量模拟结果较同化前有明显改善。2) PyWOFOST的同化模拟LAI普遍较WOFOST的模拟LAI更接近实测LAI,更符合玉米LAI的变化趋势,部分同化后的模拟LAI与实测LAI近乎重合。3)同化前WOFOST模拟发育期与实测发育期平均绝对误差为3.33天,而同化后在TSUM1的不确定性为0、10、20、30℃时PyWOFOST模拟发育期与实测发育期的平均误差分别为3.42、4.29、5.0、5.54天。4)同化后模拟得到的产量、LAI及发育期结果都充分证明了PyWOFOST模型在东北玉米种植区监测作物长势和估产的有效性,也进一步揭示了基于EnKF同化遥感观测和作物模型进行LAI模拟和作物估产的优势。5)尽管PyWOFOST的同化模拟结果较WOFOST的模拟结果普遍有改善,但并不存在所有站点在某一不确定性水平上的同化后的模拟产量或LAI全部优于其它不确定性水平的情况。6)作物模型的模拟能力直接决定同化模拟结果的优劣。由于模型本身对于严重灾害条件下作物的生长情况模拟不十分理想,因此,在严重灾害条件下,尽管同化外部观测数据后,作物的模拟产量较同化前有所改进,但仍和实测值存在较大差距。3.基于PyWOFOST的区域玉米估产研究1)以东北三省玉米种植区为例,同化区域MODIS LAI数据,模拟区域玉米产量,并利用研究区内35个地级市的统计产量验证同化模拟结果。结果表明,57.14%的区域同化估产误差在15%以内,同化产量和统计产量相关系数为0.875,确定系数为0.806；多数市的产量分布较集中,铁岭的产量标准差最小为76.16kg/ha;四平的标准差最大为1856.45 kg/ha,玉米的区域估产精度比较理想。2)作物品种遗传参数对区域作物估产影响较大,在区域作物估产过程中,应尽量将作物品种遗传参数精细化；作物制图精度也是影响作物估产精度的主要因素之一,在区域作物估产时应使用精度更高的作物种植分布图。3)总体而言,区域尺度上利用集合卡尔曼滤波同化LAI的遥感信息与作物模型模拟正常年份或轻度灾害年份作物产量的方案是可行的,但对于极端灾害条件下的作物生长模拟情况仍不十分理想。更多还原

【Abstract】 The food issue is a top priority of agricultural production. Accurate, real-time information on crop monitoring and yield forecasting is an important guarantee of the agricultural management and food security. Crop growth simulation model is a powerful tool to monitor crop growth and development, and estimate crop yield. However, there exist some difficulties in applying the crop model which is developed based on site in a wide range of region, for example, lack of crop parameters, agricultural management, and planting layout information, and not easy to solve the verification problem of model applicability on the regional scale. With the development of remote sensing technology, remote sensing has become main data support to improve crop model simulation. Crop growth model is a simplification of the actual crop growth and development process, not only model errors, there’re also errors in the initial conditions and boundary conditions of a crop model. Thus, with the forward integration of crop model, the gap between modeled value and true value will become increasingly large. Remote sensing information can display the spatial distribution of crop canopy elements, that is to provide a kind of objective observations of crop canopy elements. Therefore, it is very important to combine crop model with remote sensing observations. On the one hand, crop growth models are used to constrain remote sensing inversion model; on the other hand, remote sensing data are used to adjust the trajectory of a crop model and the accumulated errors of a crop model will get released. Different sources, spatial and temporal resolution data are integrated to express crop growth and development process on a variety of spatial and temporal scales better. Data assimilation method appears, becomes more practical and provides a feasible way for us to achieve this goal.First, considering the need for regional crop yield estimation, crop distribution patterns in the study area was studied and maize-growing area was extracted by comparing two maize mapping schemes comprehensively. Second, taking into account the problems in the domestic and foreign research and application, a foreign model (PyWOFOST) built on Ensemble Kalman Filter (EnKF) was introduced which coupled remote sensing information and crop model. First, we modified and improved the PyWOFOST model to take LAI as the joint point of the crop model (WOFOST) and remote sensing information and be applicable to the maize-growing area in Northeast China. MODIS LAI data were used as external assimilation data to simulate maize LAI, yield and development stage with the PyWOFOST model on agro-meteorological stations. Compared the results modeled by WOFOST, the results modeled by PyWOFOST which were closer to the observed values reveals the advantages of crop simulation based on data assimilation methods. At the same time, the impact of uncertainty of remote sensing observations (MODIS LAI) and model parameter (TSUM1) on the assimilation simulation results was analyzed deeply in our work. Finally, on the basis of maize mapping, the PyWOFOST model was used to estimate maize yield on a regional scale. Then the regional yield estimation result was verified by comparing with the statistical yield of maize of each prefecture-level city of three provinces in Northeast China. The result shows that the yield estimation scheme based on EnKF is applicable and feasible. The main research work and preliminary conclusions are as follows:1. Study on crop classification by remote sensing based on spectral analysis method and maize mapping.Spectral analysis method (SAM) was used to identify and classify the crop type distribu-tion in the study area, and two mapping schemes were compared comprehensively to extract maize- growing areas in the study area.1) Considering the needs to regional crop yield estimation and the limitations of measured data, MODIS products (including land cover type data, NDVI data and reflectance data) were used to extract time-series NDVI curves of various crops (soybean, maize, rice and wheat) pure pixels. And then used spectral analysis method (SAM) to identify and classify the crop type distribution in study area and extract the maize-growing area further (Mapping Scheme 1).2) The area derived from the crop classification result was compared with the crop planting area from statistical data, and the results showed that the correlation coefficient of soybeans, maize and rice were 0.858,0.715 and 0.927 respectively, and the coefficient of determination of above crops were 0.770,0.710,0.686 respectively. It revealed that the accuracy of crop classification based on spectral analysis method was ideal. In particular, for the accuracy of maize mapping, the correlation coefficient of maize inversion area extracted from crop classification result and statistical area of Heilongjiang, Jilin and Liaoning Province were 0.9527,0.6528,0.3462, with R2=0.8291,0.813,0.6883 respectively. The maize mapping accuracy of Liaoning Province was not satisfactory, so we combined 1km land cover type data from IGSNRR (Institute of Geographic Sciences and Natural Resources Research, CAS) with DEM data from NASA to extract the area with elevation range from 0-400m and dryland area accounted for more than 80% of a pixel area as the maize-growing area of Liaoning Province (Mapping Scheme 2).2. Applicability of PyWOFOST Model Based on Ensemble Kalman Filter in Simulating Maize Yield in Northeast China.The result showed that, the modeled results (mainly maize LAI, yield, development stage) after assimilating MODIS LAI were closer to observed values by comparing with the results modeled by WOFOST.1) The errors of maize production before and after assimilation at different uncertainty levels of TSUM1(0,10,20,30℃) of 20 agro-meteorological stations without the impact of meteorological disasters were 14.04%,12.71%,11.91%,10.44% and 10.48% respectively. The correlation coefficient and determination coefficient between modeled yield before assimilation and observed yield of maize was 0.681 and 0.597 respectively; the correlation coefficient between modeled yield after assimilation and observed yield of maize were all above 0.7 at different uncertainty levels of TSUM1(0,10,20,30℃)and the determination coefficient were 0.631,0.678,0.724 and 0.697 respectively. In a word, the modeled results after assimilation were better than before.2) LAI modeled by PyWOFOST which more in line with the change trend of maize LAI were generally closer to observed LAI than LAI modeled by WOFOST, part of LAI values after assimilation nearly coincided with observed LAI values.3) The mean error of development stage between modeled value of WOFOST and observed value was 3.33d; the mean error of development stage between modeled value of PyWOFOST at different uncertainty levels of TSUM1(0,10,20,30℃) and observed value were 3.42,4.29,5.0 and 5.54d respectively.4) The modeled results after assimilation (LAI, yield, development stage) all fully proved that PyWOFOT model was applicable in monitoring crop growth and estimating yield in maize-growing area in Northeast China, also revealed the advantages of assimilating remote sensing information into crop model to model LAI and estimate yield based on EnKF.5) Assimilation simulation results of PyWOFOST were better than WOFOST simulation results, but it did not exist an uncertainty level on which assimilation simulation results (yield and LAI) of all sites were superior to other uncertainty levels.6) The modeled ability of crop model directly determines the modeled result with assimilation whether good or not. The crop model itself is not very ideal for simulating crop growth and development under serious disaster conditions. Although the modeled crop yield after assimilation has improved than before, it still exist a gap between the modeled and observed value under serious disaster conditions.3. Regional maize production estimation based on PyWOFOST model1) Taking the maize-growing area in Northeast China as the study area, used MODIS LAI data as external assimilation data to estimate maize yield in the study area. Then verified the regional maize yield result based on EnKF by comparing with the statistical yield of 35 prefecture-level cities. The results showed that, the error of maize yield with assimilation were less than 15% in 57.14% of the study area; the correlation coefficient and coefficient of determination between modeled yield with assimilation and statistical yield were 0.875 and 0.806 respectively; the minimum and maximum standard deviation of modeled yield were 76.16(kg/ha) of Tieling city and 1856.45(kg/ha) of Siping city respectively. Overall, the accuracy of regional crop yield estimation was ideal.2) The genetic parameters of crop varieties has larger effect on regional crop yield estimation, it’s better to choose finer genetic parameters on regional scale; the accuracy of crop mapping was also one of the main factors for affecting the accuracy of regional crop yield estimation, consequently, more accurate crop type distribution map should be used.3) Overall, it was feasible to estimate regional crop yield under normal crop growth or mild-disaster conditions with assimilating LAI remote-sensing observations into crop model based on EnKF, but it was still not very satisfactory for simulating crop growth and estimating crop yield under serious disaster conditions.更多还原