摘要
化学需氧量(COD)是水体污染监测的常用水质指标之一,传统采集手段耗时耗力;利用遥感反演COD浓度能够快速获取整个水域的COD浓度空间分布状况,对于水污染治理和水环境保护具有重要意义。目前利用多光谱遥感数据反演COD浓度精度较低,主要原因是目前的反演模型多是以皮尔逊相关系数大小为指标选取建模波段的经验方法,对于多光谱遥感数据而言,其光谱波段范围较宽,波段的组合数量有限,难以找到有效的变量作为建模变量。针对这一问题,以郑州市天德湖为例,基于Planet多光谱高分辨率遥感影像,对遥感影像进行预处理和分析水样的高光谱数据,利用卷积神经网络对天德湖COD浓度进行反演;同时选取单变量回归模型、多变量回归模型进行精度对比。主要研究结论有:(1)相比于以皮尔逊相关系数为衡量标准选择不同波段组合的反演方式,卷积神经网络反演具有更高的空间反演精度,其决定系数为0.89,RMSE为2.22 mg·L^(-1),这是因为卷积神经网络不仅充分利用了遥感影像的光谱特征,而且能够提取目标像元周围的领域空间信息,学习到图像深层的抽象特征以及水质参数浓度和遥感数据之间的“内在规律”,可以在一定程度上避免传统方法建模带来的不稳定性;(2)选取最优的卷积神经网络模型制作天德湖水质COD浓度空间分布专题图;天德湖具有典型的内陆水体光谱特征,其COD浓度空间分布整体呈现西部高、东部较低、东南方向的进水口浓度较低、东北方向的出水口浓度较高的特征,卷积神经网络反演的天德湖区域浓度平均值为23.96 mg·L^(-1),标准差为7.11 mg·L^(-1),变异系数为0.29,更加接近实际采样点的统计值。基于卷积神经网络模型结合多光谱影像反演COD的结果表明卷积神经网络在水质参数COD遥感反演中具有较好的应用潜力。
Chemical Oxygen Demand(COD)is a commonly used water quality indicator in water pollution monitoring.Traditional collection methods are time-consuming and labor-consuming,but the inversion of COD concentration by remote sensing method can quickly obtain the spatial distribution of COD concentration in the whole water area,which is of great significance for water pollution control and water environment protection.Using multi-spectral remote sensing data inversion of COD concentration is low precision.Because at present,a lot of the inversion models based on the Pearson correlation coefficient index selection experience method,modeling band for multi-spectral remote sensing data,its wide spectral bands,and band combination of quantity is limited,hard to find effective variables as modeling.In order to solve this problem,this study in Zhengzhou city,lake as an example,based on the Planet multi-spectral high-resolution remote sensing image and the remote sensing image preprocessing and hyperspectral data for analysis of water samples,using convolution neural network method to inversion of days lake COD concentration.At the same time,choose the single variable regression model,a multivariate regression model accuracy comparison test.The main conclusions are as follows:(1)Compared with the inversion method using Pearson correlation coefficient as the measurement standard to select different band combinations,convolutional neural network inversion has higher spatial inversion accuracy,with the determination coefficient of 0.89 and RMSE of 2.22 mg·L^(-1).This is because a convolutional neural network not only makes full use of the spectral characteristics of its remote sensing images.Moreover,the spatial information of the domain around the target pixel can be extracted.The abstract features of the deep layer of the image,as well as the"internal law"between the water quality parameter concentration and remote sensing data,can be learned,which can avoid the instability caused by the traditional modeling method to a certain extent.(2)Select the optimal convolutional neural network model to make the thematic map of the spatial distribution of COD concentration in Tiande Lake water quality.Tiande Lake has typical spectral characteristics of inland water,and its spatial distribution of COD concentration is generally characterized by high in the west,low in the east,low in the southeast inlet and high in the northeast outlet.The average value of concentration in the Tiande Lake region retrieved by the convolutional neural network is 23.96 mg·L^(-1),the standard deviation is 7.11 mg·L^(-1),and the coefficient of variation is 0.29,which is closer to the statistical value of actual sampling points.The results of COD retrieval based on a convolutional neural network model and multi-spectral image show that the convolutional neural network has good application potential in remote sensing COD retrieval of water quality parameters.
作者
李爱民
范猛
秦光铎
王海隆
许有成
LI Ai-min;FAN Meng;QIN Guang-duo;WANG Hai-long;XU You-cheng(School of Geo-Science and Technology,Zhengzhou University,Zhengzhou 450001,China;School of Water Conservancy Engineering,Zhengzhou University,Zhengzhou 450001,China)
出处
《光谱学与光谱分析》
SCIE
EI
CAS
CSCD
北大核心
2023年第2期651-656,共6页
Spectroscopy and Spectral Analysis
基金
国家自然科学基金项目(U1704125)资助。
