基于HJ-1影像和环境变量的叶绿素a浓度反演——以深圳海域为例

叶绿素a的浓度能够表征海域富营养化状况，是反映海洋环境状态的一个重要指标。本研究以深圳海域为研究区，基于环境一号卫星遥感影像和实测叶绿素a浓度数据，将陆源入海排污口的核密度、距港口航道的距离和距海水增养殖区的距离作为环境变量，在Matlab平台中分别构建了基于环境一号卫星四个波段反射率为输入参数的误差逆向传播神经网络模型，及在此基础上引入了环境变量的误差逆向传播神经网络模型，以检验环境变量的引入能否提高叶绿素a浓度的反演精度，并对其输入参数进行敏感性分析。结果表明:（1）环境变量的引入能较大地提高误差逆向传播神经网络模型的反演精度，且引入环境变量的误差逆向传播神经网络模型的训练均方误差和验证均方误差分别为4.71 μg/L和3.50 μg/L，均优于原始误差逆向传播模型的10.98 μg/L和12.61 μg/L；（2）引入环境变量的误差逆向传播模型的框架如下:输入层为7个变量，分别为蓝光波段反射率、绿光波段反射率、红光波段反射率和近红外波段反射率、陆源入海排污口的核密度、距港口航道的距离和距海水增养殖区的距离；隐含层节点数为5个；输出层为叶绿素a的浓度；（3）叶绿素a的浓度对陆源入海排污口的核密度的变化最敏感，其次分别是近红外波段反射率、红光波段反射率、距港口航道距离、蓝光波段反射率、绿光波段反射率和距海水增养殖区的距离。

Retrieval of chlorophyll a concentration based on HJ-1 data and environmental variables: a case study of marine areas in Shenzhen

The concentration of chlorophyll a(Chl a) is an important indicator of eutrophication in coastal marine environments.This research retrieved the Chl a concentration of marine areas in Shenzhen using environmental variables such as the kernel density of land-based pollution outfall, distance to the harbor approach, and the distance to marine aquaculture zones.The environmental in situ data was then combined with HJ-1 multispectral data to derive two back-propagation(BP) neural-network models using Matlab.The BP neural-network models of this study were used to test whether the introduction of environmental variables could improve the accuracy of Chl a concentration estimates obtained with BP neural networks.In addition, the sensitivity of input parameters was analyzed.Results showed that:(1) The introduction of environmental variables could greatly improve the retrieval accuracy of a BP neural-network.The retrievable accuracy of the BP neural-network model modified with environmental variables was better than that of the original BP neural-network model.The MSEs(mean squared errors) of training and verification of the BP neural-network model with environmental variables were 4.71 μg/L and 3.50 μg/L, respectively.The original BP neural-network model had MSEs of training and verification of 10.98 μg/L and 12.61 μg/L, respectively.(2) The approach of using a BP neural-network model with environmental variables was investigated further.The input layer had seven variables including blue reflectance, green reflectance, red reflectance, near-infrared reflectance, the kernel density of land-based pollution outfall, distance to the harbor approach, and the distance to marine aquaculture zones.The hidden layer had five nodes.The output layer was the Chl a concentration.(3) The Chl a concentration was most sensitive to the kernel density of land-based pollution outfall, followed sequentially by the near-infrared reflectance, red reflectance, distance to the harbor approach, blue reflectance, green reflectance, and the distance to marine aquaculture zones.