【作者】 李渊；

【导师】 李云梅；

【作者基本信息】 南京师范大学 ， 遥感技术与应用， 2014， 博士

【摘要】 随着科技的不断进步,湖泊水质监测的手段也越来越丰富,这就意味着我们可以获得更多的数据源,但不同的数据具有不同的时间和空间尺度。与此同时,在国内外众多学者的不懈努力下,开发了大量的水质参数遥感估算反演模型,但不同的模型都具有其“局限性”,只能从某个层面反映“真值”。基于上述考虑,本研究以太湖为研究区,以叶绿素a浓度为研究对象,首先,利用数据同化方法,结合不同遥感估算模型的模型误差,实现太湖叶绿素a浓度的多模型协同反演；其次,利用数据同化方法,结合水体动力学模型,融合浮标、平台等观测数据及卫星影像数据,构建适合于太湖的叶绿素数据同化系统；最后,基于已构建的太湖叶绿素数据同化系统,分别以GOCI数据和浮标、平台数据为观测数据,进行了太湖叶绿素a浓度估算与预测实验。这将为多数据源、多尺度、多模型、多传感器的联合应用提供新的方法,从而克服现有的水体水质参数反演算法的不足,最终提高水质参数的反演精度。基于以上研究和分析,主要得出以下结论：(1)基于数据同化方法的多模型协同反演算法,通过对多个反演模型的反演结果进行有效加权,可以有效融合不同模型的优点,改善单模型反演精度较低区域的反演结果,从而整体提高反演精度,有利于水环境质量监测和评价。以2006～2009年太湖野外实测数据为例,最终遴选出6个反演效果较好的太湖叶绿素a浓度遥感估算模型,分别为：波段比值模型、三波段模型、四波段模型、Dall’Olmo模型、Gitelson模型和徐京萍模型,模型的平均绝对误差(MAPE)分别为29%、32%、35%、25%、27%、25%,均方根误差(RMSE)分别为13.19μg/L、14.21μg/L、28.35μg/L、14.78μg/L、13.98μg/L‘15.71μg/L、进而利用多模型协同反演算法,从6个反演模型中任意选取2至6个模型参与多模型协同反演。结果表明：当6个反演模型都参与多模型协同反演时,反演效果最好,MAPE为23.4%,RMSE为14.58μg/L；同时,随着参与多模型协同反演的模型个数的增加,反演效果也越好。(2)基于集合的卡尔曼滤波数据同化方法,可以提高太湖叶绿素a浓度的估算和预测精度。利用集合均方根滤波,结合太湖水体动力学模型,构建了太湖叶绿素数据同化系统。通过观测模拟实验(OSSE),分析和评价了太湖叶绿素数据同化系统的有效性和适用性。当浮标布设于梅梁湾时,同化实验结果较控制实验结果精度提高了65%；当浮标布设于湖心区时,同化实验结果较控制实验结果精度提高了57%；但当浮标布设于湖心区时,叶绿素a浓度分布更加连续和稳定；考虑到水质监测的需求,建议将浮标和平台布设于拖山附近。进而分别以GOCI数据和浮标、平台数据作为观测数据,利用数据同化系统进行了太湖叶绿素同化实验。当以GOCI数据为观测数据时,将同化结果、控制实验结果分别与多模型协同反演结果进行对比,全湖MAPE分别为45%和125%,决定系数分别为0.71和0.41；将同化结果、控制实验结果分别与地面实测叶绿素a浓度进行对比,MAPE分别为28%和161%,RMSE分别为9.57μg/L和55.66μg/L。当以浮标、平台数据为观测数据时,在同化阶段,对于平台站点,MAPE从218%降低到27%, RMSE从16.23μg/L降低到2.97μg/L；对于浮标站点,MAPE从1125%降低到98%,RMSE从17.29μg/L降低到3.98μg/L；在预测阶段,对于平台站点,MAPE从139%降低到3%；对于浮标站点,MAPE从2001%降低到468%。(3)太湖叶绿素数据同化系统对于不同参数的敏感性将直接影响到该系统能否准确的估算太湖叶绿素a浓度的分布。在敏感性分析实验中,分析和评价了样本数目、同化时长、背景场误差、观测误差和模型误差对于同化系统性能的影响。结果表明：从计算成本、系统运行时间和同化效果等方面分析,当集合样本数目达到30至40左右时同化系统取得了较好的结果；同化系统对于背景场误差的估计变化并不是很敏感,即初始场的估计是否准确对于同化系统影响不是很大；同化系统对于模型误差和观测误差的变化较为敏感,此外,由于水体动力学性质不一,不同的测试点位其敏感性的表现形式有所差异；利用数据同化方法可以有效的估算太湖叶绿素a浓度。更多还原

【Abstract】 With the development of technology, there are more and more ways to monitor water quality. This means that we can get more data from different sources with different time and space scales. Meanwhile, in the unremitting efforts of many scholars, large amount of remote retrieve models of water quality parameters have been developed. However, each model could only reflect the "true value" from one level because of the natural limitation of remote sensing. To get the relatively true value by combining all of the data sources with the various models, we developed the data assimilation method for retrieving the concentration of chlorophyll a in Taihu Lake: Firstly, multi-model collaborative retrieve algorithm was established using data assimilation method to retrieve chlorophyll a concentration in Taihu Lake, in which the model error of different remote retrieval models was considered to enhance the accuracy; Secondly, the chlorophyll a concentration data assimilation system was built by integrating with the water dynamics model, observation data (buoys, platforms and other observation data) and satellite imagery data; Finally, the assimilation experiments was conducted to evaluate and forecast the chlorophyll a concentration in Taihu Lake by applying the constructed data assimilation system, taking GOCI imagery data, buoys and platform data as the observation data. In this paper, a new method is provided to overcome the deficiencies of the existing remote retrieve models and ultimately improve water quality parameters retrieval accuracy, by combining multiple data sources, multi-scale, multi-model and multi-sensor.The main conclusions of this study are as follows:(1) The multi-model collaborative retrieve algorithm based on data assimilation could effectively blend the advantages of different retrieve models and meanwhile, could effectively weight the retrieve results. Therefore, it could improve the accuracy of the single model in lower retrieve accuracy region, and then improve the overall retrieval accuracy finally. In this study, six models were selected for remote retrieving chlorophyll a concentration in Taihu Lake based on in situ measured data during2006to2009, these models are:band ratio model, three band model, four band model, Dall’Olmo model, Gitelson model and Xu model. The mean absolute percent errors (MAPE) for these models are29%,32%,35%,25%,27%and25%; root mean square errors (RMSE) are13.19μg/L,14.21μg/L,28.35μg/L,14.78μg/L,13.98μg/L and 15.71μg/L. Then,2to6models were selected to test the efficiency of the multi-model collaborative retrieve processes. The results indicates that the best value is when six retrieve models all participate in the multi-model collaborative retrieve procedure, i.e., MAPE is23.4%, and RMSE is14.58μg/L. Meanwhile, with the increased retrieve model participated, the retrieve result gets better.(2) Kalman filter algorithm based on ensemble could improve the accuracy of evaluation and prediction of chlorophyll a concentration in Taihu Lake. Thereafter, the chlorophyll a data assimilation system of Taihu Lake was built using ensemble square root kalman filter, combining with water dynamic model. In addition, the effectiveness of this method for evaluation and prediction of the concentration of chlorophyll a was validated. When virtual buoys were laid in Meiliang Bay, the evaluation accuracy had been improved by65%. When virtual buoys were laid in the center of the lake, the evaluation accuracy had been improved by57%. However, the distribution of chlorophyll a is more continuous and stable when the buoys were laid in the center of the lake. Considering the water quality monitoring requirements, recommend buoys placed around Tuoshan Mountain. Then, the evaluation and forecasting of chlorophyll a concentration experiment in Taihu Lake were conducted based on Taihu Lake chlorophyll a data assimilation system, taking GOCI imagery data, buoys and platform data as the observation data.Taking GOCI imagery data as observation data, the MAPEs of assimilation experiment and control experiment compared to the multi-model retrieve result in the whole lake were45%and125%respectively, and that of the R2were0.71and0.41, respectively; While, compared to the in situ result, the MAPEs were28%and161%, respectively, and the RMSEs were9.57μg/L and55.66μg/L respectively. Taking buoys and platform data as the observation data, the MAPE for the platform station decreased from218to27%, and the RMSE decreased from16.23μg/L to2.97μg/L, during the assimilation procedure; the MAPE for the buoy station decreased from1125%to98%, and the RMSE decreased from17.29μg/L to3.98μg/L. In the prediction procedure, the MAPE for the platform station decreased from139%to3%, and the MAPE for the buoy station decreased from2001%to468%.(3) Sensibility of the Taihu Lake chlorophyll a assimilation system to different parameters directly control the accuracy of estimate the chlorophyll a concentration distribution when using this assimilation system. We used multispectral data of Environmental Satellite1(HJ-1), obtained on21April,2009, combined with in situ data to retrieve the concentration of chlorophyll a in Taihu Lake. Take the retrieved chlorophyll a concentration of Taihu Lake as the initial background value, then combined with the data assimilation system to analyze the influence of the ensemble size, the assimilation time, the background error, the observation error and the model error on the assimilation system. The results indicate:taking the computing cost, time cost of system and the performance of assimilation system into consideration, the assimilation system performs well when the ensemble size are30-40; the assimilation system is not very sensitive to the background error; both the observation error and the model error are very sensitive for the performance of the system; different test stations have different water dynamic properties, that induces the different performance; the estimation of chlorophyll a concentration can be improved by using the data assimilation method.更多还原