梁益同, 胡江林. NOAA卫星图像神经网络分类方法的探讨[J]. 武汉大学学报 ( 信息科学版), 2000, 25(2): 148-152.
引用本文: 梁益同, 胡江林. NOAA卫星图像神经网络分类方法的探讨[J]. 武汉大学学报 ( 信息科学版), 2000, 25(2): 148-152.
LIANG Yitong, HU Jianglin. Application of Neural Network Technique to Classification of NOAA Satellite Images[J]. Geomatics and Information Science of Wuhan University, 2000, 25(2): 148-152.
Citation: LIANG Yitong, HU Jianglin. Application of Neural Network Technique to Classification of NOAA Satellite Images[J]. Geomatics and Information Science of Wuhan University, 2000, 25(2): 148-152.

NOAA卫星图像神经网络分类方法的探讨

Application of Neural Network Technique to Classification of NOAA Satellite Images

  • 摘要: 阐述了应用神经网络对NOAA卫星图像进行分类的基本原理和方法,并进行了实例分析,结果表明了该方法的有效性。

     

    Abstract: Classification of NOAA satellite image refers to partitioning the information of NOAA satellite images into different types of targets for this information to be extracted easily,which is the basis of using NOAA satellite data.The traditional approaches of classification are visual inter-pretation and threshold method.In the visual interpretation,experienced interpretation is needed and there is always unfavorableness in the description of information distribution and the improvement of efficiency.The threshold method only pays attention to the spectrum of one type of target at some channels and ignores it at other channels.Actually the character of the spectrum of this type is different from other types.In addition,it is difficult that many thresholds are adjusted at the same time to appropriate value.Due to these accounts,the precision of the threshold method can not be improved,and it can not be used becomingly because the channel numbers of the detector of satellite will be increased in the future.The artificial neural network technique developing rapidly in recent years provides a new means for ameliorating this problem.Neural network has the basic characters of human brains,such as learning,recollection and generalizing.The peculiarity of neural network is massive parallel computing,distributing memory of information,nonlinear dynamics of consecutive time,global behaviour,great fault-tolerance and robust,self-organization,self-learning and real time processing. In this paper,the principles and method of using artificial neural network technique to automatically classify NOAA satellite image are discussed and the analysis of NOAA/AVHRR data on August 4,1998 is presented.The method can be briefly described as follows: ①Classify the information of the images into three types of targets: water body,cloud and land. ②Normalize the data of all channels of NOAA satellite.③Create a neural network of BP model having input layer,hidden layer and output layer with neuron numbers of 4,5 and 3.④Pick up a few image elements of different types of targets as trained samples by the visual interpretation.In this paper,the sample numbers of water body,cloud and land are 78,65 and 49 respectively.⑤Regard real type of the samples as the output of the network and the corresponding channels data as the input,train the network until its cost function are steady by adjusting the weights and thresholds between interconnecting neurons.⑥Input the channel data of NOAA satellite into the disciplined network and judge the type of target of all image elements according to the output of network.The result of application shows that the disciplined network has recollected the characters of spectrum of all types of targets at all channels.The network is provided with the power to classify NOAA satellite images and the precisions of water body,cloud and land are 96%,92% and 87.5% respectively and the average precision is 91.8%.In this paper,two conclusions are drawn,one is that neural network can be used for classification of NOAA satellite images and better precision is obtained,the other is that neural network uses the experience of visual interpretation for reference,as well as pays adequately attention to the spectrum of one type of target at all channels,so it's an important approach of classification of NOAA satellite images.

     

/

返回文章
返回