【作者】 顾万花；

【导师】 马蔚纯；

【作者基本信息】 复旦大学 ， 环境科学， 2011， 硕士

【摘要】 湖泊是重要的淡水资源,但是随着经济的高速发展和工业化程度的加剧以及人为活动的影响,湖泊污染和富营养化问题日益严重,加强湖泊水质的监测和治理非常有必要。常规的水质监测耗时耗力,并且很难反映湖泊的整体水质状况,而遥感技术可以快速、及时的提供整个湖区的水质状况,具有监测范围广、成本低和便于进行长期动态监测的特殊优势,在湖泊水质监测中具有巨大的应用潜力。内陆水体的叶绿素a浓度是水质评价重要指标之一。本文在总结分析国内外叶绿素a浓度遥感估算已有的理论、技术和方法的基础上,运用半经验模型和人工神经网络,利用2010年8月23、9月7日和9月8日的野外实测数据、实测光谱数据和环境1号星、MODIS卫星遥感影像,探讨了淀山湖水体叶绿素a浓度的估算模型。主要研究内容和结论如下：1)基于实测光谱数据,构建了叶绿素a的半分析模型。通过迭代,选取三个波段,以[Rrs-1(663)-Rrs-1(689)]×Rrs(759)为自变量构建模型,线性和曲线模型的R2分别为0.519和0.87。2)针对不同卫星数据源构建半经验析模型。环境1号星高光谱数据的三波段组合[Rrs-1(70)-Rrs-1(77)]×Rrs(88)所建立的线性和曲线模型的R2分别为0.934和0.955,进行对叶绿素a的监测是可行的。但是无论是环境1号星还是MODIS卫星,基于多光谱数据的半经验模型用于内陆湖泊水质研究还有一定的困难。3)基于人工神经网络建立估算叶绿素a浓度的模型。高光谱数据三波段输入的神经网络模型的决定系数R2最大为0.938。基于环境1号星高光谱数据、多光谱数据和MODIS的人工神经网络模型的R2最大为0.992、0.87和0.91。对于不同的遥感数据源,人工神经网络模型的精度均优于半分析模型。表明人工神经网络更加有利于反映和提取淀山湖水体的基本光学特征,在光谱特征复杂的内陆水体水质参数估测中具有优越性。4)将溶解氧、化学需氧量、总磷、总氮也作为人工神经网络模型的输入值,构建基于实测光谱数据的改进的反演叶绿素a浓度的人工神经网络模型的R2最大为0.96,有效提高了反演精度。论文构建了估算淀山湖叶绿素a浓度的模型,高光谱数据模型精度较高,优于卫星多光谱数据。人工神经网络有效提高了估算精度,证明神经网络方法在复杂的内陆水体水质反演中具有较大的优越性,且在网络输入中加入溶解氧、化学需氧量、总磷、总氮因子,为叶绿素a的遥感反演提供了新的思路。更多还原

【Abstract】 Lakes are very important in freshwater supply. At present, most of the lakes in the world are suffering from eutrophication with industrialization, economic growth, and human activities. Frequent monitoring and integrated management of lakes are neeessary. Howeve, conventional methods for monitoring in lakes which are time& money-consumed can’t provide water quality information of the whole lake. With spatial and temporal information, remote sensing technique is suitable to obtain such information quickly and frequently, and it’11 be widely applied in water quality monitoring of lakes in future because of its advantage of fastness, economy and long-term dynamic monitoring.Chlorophlly-a(chl-a) is one of the most important indexs for estimating inland water quality. Based on the analysis of former theory, techniques and methods in eastimating chl-a concentration, this paper chooses semiempirical model and artificial neural network, using in situ measurements of chl-a concentration on Augest 27th, September 7th and 8th, the synchronous, HJ-1 and MODIS imageries and then discusses remote sensing eastimation of chl-a eastimating chl-a concentration in Dianshanhu Lake. The major contents and conclusions are introduced as following:1) Study on semiempirical model for estimating chl-a concentration based on spectrum data:choosing three bands to make [Rrs-1663)—Rrs-1(689)]×Rrs(759) as independent variable and build lined and curve models, which R2 are 0.519 and 0.87.2) The stusy develops semiempirical models based on different satellites data. To hyper spectral data of HJ-1, R2 of lined and curve models with [Rrs-1(70)—Rrs-1(77)]×Rrs(88) as independent variable are 0.934 and 0.955. To multi-spectral data, such as HJ-1 and MODIS, there lies difficulty on the application of semiempirical models.3) Study on estimation model based on artificial neural network. To spectrum data, the maximam of R2 of artificial neural network model is 0.938, whose input is three bands, while that of hyperspectral data and multi-spectral data of HJ-1 and MODIS are 0.992,0.87 and 0.91. The artificial neural network models’ estimation precision is higher than regression models’.4) To improve the estimation precision of artificial neural network model, adding DO, COD, TP and TN into input data. As a result R2 is elevated to 0.96. This paper develops models for estimating chl-a concentration in Dianshanhu Lake, acquires good estimation precision. Hyper spectral data is batter than multi-spectral data. It is proved that the artificial neural network model for estimating water quality of complex inland water body possesses advantage. Inputing DO, COD, TP and TN factors into the neural network provides a new way for estimating chl-a concentration.更多还原