文献类型：期刊文献

中文题名：基于LSTM网络的海水石油污染含量遥感预测模型

英文题名：Prediction Model of Petroleum Pollution Content in Seawater Based on LSTM Network and Remote Sensing

作者：黄妙芬[1];王江颖[1];邢旭峰[1];王忠林[1];周运[1]

机构：[1]广东海洋大学数学与计算机学院,广东湛江524088

年份：2021

卷号：41

期号：5

起止页码：67

中文期刊名：广东海洋大学学报

外文期刊名：Journal of Guangdong Ocean University

收录：CSTPCD、、北大核心、北大核心2020

基金：国家自然科学基金项目(41771384);国家重点研发计划重点专项资助项目(2016YFC1401203)。

语种：中文

中文关键词：陆地卫星Landsat;长短期记忆网络(LSTM);遥感反射比;石油污染含量;预测模型

外文关键词：Landsat;Long Short-Term Memory network(LSTM);remote sensing reflectance;petroleum pollutant content;prediction model

中文摘要：【目的】建立一种基于美国陆地卫星Landsat遥感数据和长短期记忆(Long Short-Term Memory,LSTM)网络的海洋石油污染含量预测模型。【方法】利用1984-2020年在大连新港海域过境的Landsat卫星所采集的可见光-近红外波段(0.45~0.90μm)光谱数据,基于LSTM网络,分别建立空间分辨率为30 m、时间分辨率为8 d的4波段遥感反射比Rrs预测模型,并对预测模型所涉及的神经网络层数、隐藏神经元节点数和回溯时间步长等超参数进行优化;在4波段Rrs预测值的基础上,结合基于水体石油污染归一化遥感反射比指数(normalized difference petroleum remote sensing reflectance index,NDPRI)的石油含量遥感反演模型,对海域石油污染含量Co值进行预测。【结果】对于蓝光、绿光、红光和近红外4个波段,神经网络层数依次取3、3、4和3层,隐藏神经元节点取64、96、64和96个,回溯时间步长皆取6 d为最优;根据2021年1-5月现场的Co测量值,对LSTM网络预测值进行精度分析,平均相对误差为9.17%。【结论】基于LSTM网络建立的Co预测模型具有较好的精度,所预测的结果可弥补在有云情况Co数据缺失的问题,也可为相关Co未来动态演变研究提供一种新技术手段。

外文摘要：【Objective】A prediction model was proposed using the Long Short-Term Memory network(LSTM)with Landsat images to estimate the petroleum pollutant content in sea water.【Method】The model employed the LSTM network and visible and near-infrared(0.45-0.9μm)Landsat images in Dalian Xingang coastal area were retrieved from 1984 to 2020.Those images were used to construct a 4-band remote sensing reflectance(Rrs)prediction model with spatial resolution of 30m and temporal resolution of 8 days.Then,the 4-band Rrs prediction result was used together with the normalized difference in petroleum remote sensing reflectance index(NDPRI)based petroleum content inversion model and,to predict the petroleum content(Co)in the specific sea area.【Result】The prediction model achieved its best performance when the numbers of neural network layers at blue,green,red and near-infrared bands were set as 3,3,4 and 3 respectively;the hidden neuron nodes were 64,96,64,and 96 respectively.Also,the backtracking time step was set at 6 days.Compared with the field data of Cocollected between January and May of 2021,the precision analysis of LSTM showed a mean relative deviation of 9.17%between the field data and the predicted values given by the model.【Conclusion】The LSTM network can predict relatively high accuracy in Co.The predicted results not only can fill the gap when Co data is missing due to unavailability of Landsat data in the cloudy days,but also provide a new approach to study the dynamics of petroleum pollutant content over time in this kind of contaminated sea area.