基于Sentinel-2的平寨水库叶绿素a浓度反演

为实现对平寨水库叶绿素a的遥感监测，选取平寨水库2017年11月17—18日的实测叶绿素a浓度数据和准同步的Sentinel-2数据，通过选取最佳波段组合建立BP神经网络模型，对平寨水库叶绿素a进行反演，并分析其空间分布特征。结果表明:Sentinel-2红边波段对叶绿素a的敏感性优于可见光波段，在叶绿素a浓度反演方面具有较大潜力。相关系数最大的波段组合方式是:B5/B4、1/B4-1/B5]*B6、1/B4-1/B5]*B7和1/B4-1/B5]*B8；BP神经网络模型可决系数R2为0.9160，平均相对误差为29.87%，反演精度优于三波段模型；平寨水库叶绿素a浓度空间分布差异明显，水面开阔的中心库区浓度较高，各支流上游河段浓度较低。Sentinel-2数据可较好地应用于喀斯特高原湖泊叶绿素a浓度反演，BP神经网络模型估测结果合理、可靠；研究结果可为平寨水库水环境治理提供科学依据。

RETRIEVAL OF CHLOROPHYLL-A CONCENTRATION IN PINGZHAI RESERVOIR BASED ON SENTINEL-2

To realize remote sensing monitoring of chlorophyll-a in Pingzhai Reservoir, the measured chlorophyll-a concentration and quasi-synchronized Sentinel-2 data of Pingzhai Reservoir on November 17th and 18th, 2017 were selected. The BP neural network model was established by selecting the best band combination to invert the chlorophyll-a of Pingzhai Reservoir, and its spatial distribution characteristics was analyzed. The Sentinel-2 red edge band was more sensitive to chlorophyll-a than the visible light band and had greater potential for chlorophyll-a concentration inversion. The band combination method with the largest correlation coefficient were: B5/B4, 1/B4-1/B5]*B6, 1/B4-1/B5]*B7, and 1/B4-1/B5]*B8; the resolvable coefficient R2 of BP neural network model was 0.9160 and the average relative error was 29.87%. The inversion accuracy of BP neural network model was better than that of three-band model; the concentration distribution of chlorophyll-a in Pingzhai Reservoir was obviously different. The concentration of the central reservoir in the open water was higher, and the concentration in the upper reaches of each tributary was lower. The research showed that Sentinel-2 data could be well applied to the retrieval of chlorophyll-a concentration in karst plateau lakes. The prediction results of BP neural network model was reasonable and reliable. The research results could provide a scientific basis for the water environment management of Pingzhai Reservoir.