基于BP神经网络的鄱阳湖叶绿素a遥感反演

Remote sensing inversion of chlorophyll-a in Poyang Lake based on BP neural network

摘要

【目的】通过对比两种模型对鄱阳湖叶绿素a的反演结果,筛选出反演精度更高的一种方法,使其能够更准确、高效地应用于浅水湖泊的水质监测与管理。【方法】叶绿素a是水环境监测的重要指标,是衡量水体富营养化的关键参数。本文以鄱阳湖为典型研究区域,基于实测叶绿素a浓度和Landsat-8 OLI卫星遥感数据,对湖区叶绿素a浓度进行反演。通过对遥感影像的预处理,利用单波段及波段组合数据与叶绿素a浓度数据进行相关性分析,构建了叶绿素a波段比值模型,筛选相关波段组合,在此基础上进一步构建了BP神经网络模型(Back Propagation Neural Network Model),将实测叶绿素a浓度与BP神经网络模型反演结果进行相关性比较。【结果】结果表明:所构建的BP神经网络模型相较于波段比值模型预测值与实测值之间的决定系数(R2)从0.624~0.855提升到0.745~0.921,平均绝对误差百分比(MAPE)和均方根误差(RMSE)相较于波段比值模型降低46%以上。【结论】BP神经网络模型在反演精度上优于波段比值模型,时间尺度上,基于BP神经网络模型反演的丰水期叶绿素a浓度高,枯水期低,夏季叶绿素a浓度升高,冬季降低;空间尺度上,湖心及水体流动强的区域叶绿素a浓度低,沿岸及人类活动强的区域浓度高,南部湖区高于北部。本研究所构建的BP神经网络模型对浅水湖泊叶绿素a反演效果较好,对湖泊生态环境保护具有重要支撑。

Abstract

[Objective] By comparing the inversion result of chlorophyll-a from two models in Poyang Lake, the model with higher inversion accuracy is selected, enabling more accurate and efficient application in water quality monitoring and management of shallow lakes. [Methods] Chlorophyll-a is a key indicator for water quality monitoring and a critical parameter for eutrophication assessment in aquatic environments. Poyang Lake was selected as a representative study area. Chlorophyll-a concentration in Poyang Lake was inverted based on measured chlorophyll-a concentration and Landsat-8 OLI satellite remote sensing data. After preprocessing the remote sensing images, correlation analysis was conducted between single band and band combination data and chlorophyll-a concentration data. A chlorophyll-a band ratio model was developed, and relevant band combinations were selected to further establish a back propagation( BP) neural network model. The correlation between the measured chlorophyll-a concentrations and the inversion result from the BP neural network model was compared. [Results] The result showed that the developed BP neural network model led to an improvement in the coefficient of determination(R2) between the predicted and measured values, from 0. 624~0. 855 to 0. 745~0. 921, compared to the band ratio model. The mean absolute percentage error(MAPE) and root mean square error(RMSE) were reduced by more than 46% compared to the band ratio model. [Conclusion] The BP neural network model outperforms the band ratio model in inversion accuracy. Temporally, chlorophyll-a concentrations inverted by the BP neural network model are higher during wet seasons and lower during dry seasons, with chlorophyll-a concentrations increasing in summer and decreasing in winter. Spatially, chlorophyll-a concentration is lower in the central lake and areas with high water flow, and higher along the shoreline and in regions with intense human activities, with the southern lake area showing higher concentrations than the northern area. The established BP neural network model demonstrates excellent performance in chlorophyll-a inversion in shallow lakes, providing important support for the conservation of ecological environment in lakes.

关键词

遥感影像 / 叶绿素a / 神经网络 / 浅水湖泊 / 影响因素

Key words

remote sensing image / chlorophyll-a / neural network / shallow lakes / influencing factors

基金资助

国家重点研发计划项目(2022YFC3202005)

国家自然科学基金项目(52209086)

中国空间技术研究院CAST创新基金项目