基于EnMAP卫星和深度神经网络的LAI遥感反演方法
投稿时间:2019-03-23  修订日期:2019-09-11  点此下载全文
引用本文:
摘要点击次数: 63
全文下载次数: 1
作者单位E-mail
李雪玲 首都师范大学 数学科学学院,北京 100048
中国科学院遥感与数字地球研究所数字地球重点实验室北京100094
中国科学院空天信息研究院北京100094 
515731746@qq.com 
董莹莹 中国科学院遥感与数字地球研究所数字地球重点实验室北京100094
中国科学院空天信息研究院北京100094 
dongyy@radi.ac.cn 
朱溢佞 首都师范大学 数学科学学院,北京 100048  
黄文江 中国科学院遥感与数字地球研究所数字地球重点实验室北京100094
中国科学院空天信息研究院北京100094 
 
基金项目:国家自然科学基金“结合形态和营养指标的小麦长势遥感监测方法” (41601466);国家重点研发计划项目 “粮食作物重大病虫害遥感监测预警与防控技术” (2017YFE0122400);中国科学院青年创新促进会 (2017085);北京市教委科技计划一般项目(KM201710028002)
中文摘要:区域叶面积指数(Leaf Area Index, LAI)定量反演是开展大尺度农作物长势监测和产量估算的重要基础。针对当前区域LAI遥感定量反演存在的反演精度不理想和模型稳定性弱等问题,提出了一种基于少量训练样本进行LAI高精度反演的深度神经网络(Small Simple Learning LAI-Net, SSLLAI-Net)。该网络由2个卷积层、1个池化层和3个全连接层构成,将光谱反射率数据作为网络输入端,输出端得到LAI反演值,且该网络模型可支持小样本数据量的训练。本文以德国阿尔卑斯山麓高光谱遥感卫星影像Environmental Mapping and Analysis Program (EnMAP)为数据源,以该区域的谷物、玉米、油菜、其他农作物为研究对象,数值实验结果表明当各作物类别的训练样本量均为50时,基于SSLLAI-Net的LAI反演精度分别为0.95、0.99、0.98、0.90;且在添加噪声的情况下,各作物类别的LAI反演精度分别为0.95、0.98、0.96、0.89。综上,本文提出的基于深度神经网络的区域LAI遥感定量反演方法SSLLAI-Net是鲁棒可靠的,且该模型能够支持稳定的小样本建模。
中文关键词:叶面积指数  高光谱遥感  EnMAP  深度神经网络  SSLLAI-Net  
 
Leaf area index estimation with EnMAP hyperspectral data based on deep neural network
Abstract:Regional leaf area index (LAI) mapping is important for crop growth monitoring and yield estimation. Due to the lower accuracy and instability of statistical models for regional LAI estimation, we proposed a new deep neural network model, i.e. Small Simple Learning LAI-Net (SSLLAI-Net), based on small sample training, to achieve stable relationship between hyperspectral reflectance and LAI. The new proposed SSLLAI-Net was constructed with two convolution layers, one pooling layer and three connect layers, for which the inputs and outputs were hyperspectral reflectance and LAI estimation. Moreover, SSLLAI-Net could support small training sets. We applied SSLLAI-Net to an Environmental Mapping and Analysis Program (EnMAP) hyperspectral imagery for regional LAI mapping, in which cereals, maize, rape seed and other crops are selected as our objects. The achieved R2 values for estimated LAI of cereals, maize, rape seed and other crops were 0.95, 0.99, 0.98 and 0.90 based on small training sets with 50 samples, while for the inputs with noise, the R2 values were 0.95、0.98、0.96 and 0.89, respectively. In all, our new proposed SSLLAI-Net has high precision of regional LAI mapping, stability and noise resistance with hyperspectral remote sensing observations.
keywords:leaf area index (LAI), hyperspectral remote sensing, EnMAP, deep neural network, SSLLAI-Net
  HTML  查看/发表评论  下载PDF阅读器

版权所有:《红外与毫米波学报》编辑部