基于Sentinel-2影像的洱海叶绿素a质量浓度反演

为动态监测洱海水体富营养污染物，利用遥感技术对反映水体富营养化的核心参数——叶绿素a质量浓度进行反演，建立适合当地当季的反演模型，对水体叶绿素a质量浓度进行宏观监测；通过洱海的秋季Sentinel-2影像和实测叶绿素a质量浓度数据，使用参数相关分析方法选取反演波段，建立BP神经网络模型和多元线性回归模型，随机选择7个样本点对2种模型进行交叉验证后，对洱海叶绿素a质量浓度进行反演。结果表明：Sentinel-2数据与叶绿素a质量浓度具有显著的相关关系(Pearson积矩相关系数的绝对值大于0.7, P < 0.001)，且分别在单波段、单波段比值和双波段比值中相关系数最大的波段及波段组合为B6、B7/B6和(B6+B8)/(B7+B8a)；隐含层神经元节点数为4的3层BP神经网络模型的均方根误差最小，决定系数最大，分别为0.002 8和0.925；2019年10月12日、11月9日，洱海叶绿素a质量浓度在空间上均呈北部高于南部的分布状态；BP神经网络模型的平均绝对误差百分比为21.36%、均方根误差为0.002 8、决定系数为0.925，多元线性回归模型的平均绝对误差百分比为27.90%、均方根误差为0.004 5、决定系数为0.788。总体而言，BP神经网络模型的叶绿素a质量浓度反演精度高于多元线性回归模型。本研究成果可为相关部门对洱海水质进行动态监测以及制定洱海水质保护措施提供参考。

Mass concentration inversion for chlorophyll a in Erhai lake based on Sentinel-2

In order to dynamically monitor eutrophic pollutants in Erhai lake, the remote sensing technology was used to invert the chlorophyll-a mass concentration, the core parameter reflecting eutrophication of water. The inversion model suitable for the local season was established to conduct the macro monitoring of the mass concentration of chlorophyll-a in water. Based on Sentinel-2 images and measured mass concentration data of chlorophyll a in Erhai lake in autumn, the inversion bands were selected by parameter correlation analysis method, then BP neural network model and multiple linear regression model were established. Seven sample points were randomly selected to cross-verify the two models, and then the mass concentration of chlorophyll a in Erhai lake was inverted. The results showed that a significant correlation occurred between Sentinel - 2 data and the mass concentration of chlorophyll a (the absolute value of Pearson's product moment correlation coefficient was higher than 0.7, P < 0.001), and the bands or band combinations with the largest correlation coefficient in single band, single band ratio and dual band ratio were B6, B7 / B6 and (B6 + B8) / (B7 + B8a), respectively; the three-layer BP neural network model with four neuron nodes in hidden layer had the smallest root mean square error and the largest determination coefficient, which were 0.0028 and 0.925, respectively. On October 12th and November 9th, 2019, the spatial distribution of mass concentration of chlorophyll a in the northern part of Erhai lake was higher than that in the southern part. The mean absolute percentage error of BP neural network model was 21.36%, the root mean square error was 0.002 8, and the coefficient of determination was 0.925. The mean absolute percentage error of multiple linear regression model was 27.90%, the root mean square error was 0.004 5, and the coefficient of determination was 0.788. In general, the inversion accuracy of mass concentration of chlorophyll-a by BP neural network model was higher than that by multiple linear regression model. The results of this study can provide a reference for relevant departments to dynamically monitor water quality of Erhai lake and formulate water quality protection measures of Erhai lake.