基于BP神经网络遥感图像特征分类方法的研究
|
摘要
遥感技术是一种能够对目标进行非接触测量,采集并且分析的一种新型探测技术。随着电子计算机和空间技术的发展,各种资源和环境卫星的发射和成功的运行,通过遥感卫星从太空的高度对地球全貌以及地表动态变化等各种资源信息的提取技术得到了快速的发展。因此,对获得的海量遥感图像进行识别处理,即通过提取图像信息的特征,并利用这些特征进行图像分类,进而达到图像识别一直是遥感技术所要解决的重要问题之一。
神经网络因特有的自组织学习和强大的容错性,并且在解决非线性映射问题上表现出独特的功能,使得神经网络比传统的统计参数分类方法上表现出很大的优势,还能有一定的抗噪能力,被广泛的应用在模式识别以及图像处理等各种领域。针对已经有用于遥感图像处理的神经网络系统,本文对其应用最广泛的BP神经网络进行了介绍。由于标准的BP算法在对图像分类时存在着收敛速度慢以及对网络初始参数的依赖,容易陷入局部最小值问题。针对BP算法的缺点,我们利用具有全局搜索能力、强继承性的遗传算法对BP初始网络参数进行训练,得到的最优染色体可以提高分类效率和解决局部极值的问题。因此本文构造了基于GA-BP的三层神经网络用于遥感图像进行分类的方法。
一般在处理遥感图像分类问题时采用的是基于光谱特征的分类方法,即针对遥感图像的灰度特征进行分类,采用单特征进行分类的精度不高,所以本文采用的是基于光谱和纹理特征相融合的分类方法。对于光谱特征的提取,采用的是对整幅图像进行加窗提取窗口内像素的平均灰度特征的方法,并用提取的平均灰度特征来表示该窗口区域的特征值。对于纹理特征提取也是采用同样的方法,先求出窗口内像素的灰度共生矩阵,然后求平均相关性特征和能量特征,用这三个特征来代表整个窗口内图像像素的特征信息,将这三个特征值进行标准化处理作为GA-BP网络的数据输入,然后分别对采用灰度特征值和融合特征值作为输入的图像进行分类,并且进行评价。结果表明采用改进的GA-BP网络对融合特征值的网络输入,分类结果的精度和分类速度都高于对单特征的分类效果。
Remote sensing technology is new scientific detection technology. It is a realization of the goal of long-range, non-contact measurement, target acquisition and analysis of a technology. Along with the computer and the development of space technology, all kinds of resources and environment the satellite launch and successful operation, through satellite remote sensing the height of the earth from space to panorama and the dynamic changes of various resources such as surface information extraction technology obtained fast development. Therefore, the remote sensing image recognition processing, namely through the extraction of image feature information, and use these features for image classification, to achieve image recognition has been one of the important problems to solve.
Neural network for special organization learning and strong from the fault tolerance, and to solve nonlinear mapping problems in liquor unique function.Neural network than traditional statistical parameters classification method borrows a lot of advantage, still can have certain anti-noise ability, by widespread application in pattern recognition and image processing, and etc. According to have used in remote sensing image processing of neural network system, this paper on the application of the most extensive BP neural network are introduced in this paper. Due to the standard BP algorithm in the slow rate of convergence and the dependence of the initial parameters of the network image classification, it is easy to fall into the local minimum value problems. According to the weaknesses of the BP algorithm, we use the algorithm of the global search ability and strong hereditary genetic to train the BP initial network parameters, not only can Get the optimal chromosomes and improve classification efficiency, but also can solve the local minimum problem. Therefore this paper constructs a method, for remote sensing which is based on GA-BP three-layer neural network.
Generally used when dealing with remote sensing image classification is a classification method based on spectral characteristics, namely, gray feature for remote sensing image classification, using the single-feature classification accuracy is not high, so this paper is based on spectral and texture features integration of the classification. To the extraction of the spectrum characteristics, using the method of extracting average gray feature. Also use the same method for texture feature extraction, first calculate the GLCM of pixels in the window, and then calculate the average correlation characteristics and energy characteristic,and using three characteristics to represent the characteristic information of the image pixels within the entire window. After that, we make these three eigenvalues be normalized as the GA-BP network data input; and we classify and evaluate the image which make use of gray feature value and integration of the characteristic value as the input. The results show that using the improved GA-BP network integration of the eigenvalue of the network input, the classification accuracy and classification speed is higher than the classification results of single-feature. Figure24table15reference44
神经网络因特有的自组织学习和强大的容错性,并且在解决非线性映射问题上表现出独特的功能,使得神经网络比传统的统计参数分类方法上表现出很大的优势,还能有一定的抗噪能力,被广泛的应用在模式识别以及图像处理等各种领域。针对已经有用于遥感图像处理的神经网络系统,本文对其应用最广泛的BP神经网络进行了介绍。由于标准的BP算法在对图像分类时存在着收敛速度慢以及对网络初始参数的依赖,容易陷入局部最小值问题。针对BP算法的缺点,我们利用具有全局搜索能力、强继承性的遗传算法对BP初始网络参数进行训练,得到的最优染色体可以提高分类效率和解决局部极值的问题。因此本文构造了基于GA-BP的三层神经网络用于遥感图像进行分类的方法。
一般在处理遥感图像分类问题时采用的是基于光谱特征的分类方法,即针对遥感图像的灰度特征进行分类,采用单特征进行分类的精度不高,所以本文采用的是基于光谱和纹理特征相融合的分类方法。对于光谱特征的提取,采用的是对整幅图像进行加窗提取窗口内像素的平均灰度特征的方法,并用提取的平均灰度特征来表示该窗口区域的特征值。对于纹理特征提取也是采用同样的方法,先求出窗口内像素的灰度共生矩阵,然后求平均相关性特征和能量特征,用这三个特征来代表整个窗口内图像像素的特征信息,将这三个特征值进行标准化处理作为GA-BP网络的数据输入,然后分别对采用灰度特征值和融合特征值作为输入的图像进行分类,并且进行评价。结果表明采用改进的GA-BP网络对融合特征值的网络输入,分类结果的精度和分类速度都高于对单特征的分类效果。
Remote sensing technology is new scientific detection technology. It is a realization of the goal of long-range, non-contact measurement, target acquisition and analysis of a technology. Along with the computer and the development of space technology, all kinds of resources and environment the satellite launch and successful operation, through satellite remote sensing the height of the earth from space to panorama and the dynamic changes of various resources such as surface information extraction technology obtained fast development. Therefore, the remote sensing image recognition processing, namely through the extraction of image feature information, and use these features for image classification, to achieve image recognition has been one of the important problems to solve.
Neural network for special organization learning and strong from the fault tolerance, and to solve nonlinear mapping problems in liquor unique function.Neural network than traditional statistical parameters classification method borrows a lot of advantage, still can have certain anti-noise ability, by widespread application in pattern recognition and image processing, and etc. According to have used in remote sensing image processing of neural network system, this paper on the application of the most extensive BP neural network are introduced in this paper. Due to the standard BP algorithm in the slow rate of convergence and the dependence of the initial parameters of the network image classification, it is easy to fall into the local minimum value problems. According to the weaknesses of the BP algorithm, we use the algorithm of the global search ability and strong hereditary genetic to train the BP initial network parameters, not only can Get the optimal chromosomes and improve classification efficiency, but also can solve the local minimum problem. Therefore this paper constructs a method, for remote sensing which is based on GA-BP three-layer neural network.
Generally used when dealing with remote sensing image classification is a classification method based on spectral characteristics, namely, gray feature for remote sensing image classification, using the single-feature classification accuracy is not high, so this paper is based on spectral and texture features integration of the classification. To the extraction of the spectrum characteristics, using the method of extracting average gray feature. Also use the same method for texture feature extraction, first calculate the GLCM of pixels in the window, and then calculate the average correlation characteristics and energy characteristic,and using three characteristics to represent the characteristic information of the image pixels within the entire window. After that, we make these three eigenvalues be normalized as the GA-BP network data input; and we classify and evaluate the image which make use of gray feature value and integration of the characteristic value as the input. The results show that using the improved GA-BP network integration of the eigenvalue of the network input, the classification accuracy and classification speed is higher than the classification results of single-feature. Figure24table15reference44
引文
[1]戴昌达,姜小光,唐伶俐等.遥感图像处理与分析[M].北京：清华大学出版社,2004.
[2]倪金生,蒋一军,张富民.遥感图像处理理论与实践[M].北京：电子工业出版社,2008.
[3]朱述龙,朱宝山,王红卫.遥感图像处理与应用[M].北京：科学出版社,2002.
[4]王爱玲,叶明生,邓秋香MATLAB R2007图像处理技术与应用[M].电子工业出版社,2008.
[5]李爽,丁圣彦,许叔明.遥感影像分类方法比较研究fJ].河南大学学报,2002(32)：70-73.
[6]孙家炳.遥感原理与应用[M].第1版.武汉：武汉大学出版社,2003.
[7]魏海坤.神经网络结构设计的理论与方法[M].国防工业出版社,2005.
[8]王翔.基于神经BP神经网络的遥感影像模式识别方法研究[D].太原：太远科技大学,2009.
[9]甘井中.基于BP神经网络图像识别的研究[D].大连：大连理工大学,2008.
[10]李登科,张京红,戴进,遥感图像处理系统ERDAS IMAGINE及其应用[J].陕西气象,2002,(2)：20-22.
[11]刘宣江.基于神经网络的遥感影像识别[D].北京：北京邮电大学,2007.
[12]刘安斐,李弼程,张先飞.基于数据融合的多特征遥感图像分类[J].数据采集与处理,2006,21(4)：463-467.
[13]安斌,陈书海等.纹理特征在多光谱图像分类中的应用[J].激光与红外,2002,32(3)：188-190.
[14]孙艳霞.纹理分析在遥感图像识别中的应用[D].新疆：新疆大学,2005.
[15]徐青张艳等.遥感影像融合与分辨率增强技术[M].北京：科学出版社,2007
[16]王知鸷.基于基于纹理及光谱信息融合的遥感图像分类方法研究[D].西安：西安科技大学,2010.
[17]杨听,汤国安等.ERDAS遥感数字图像处理实验教材[M].北京：科学出版社,2009.
[18]张海涛.光谱与纹理特征融合的遥感图像分类方法[D].西安：西安电子科技大学,2010.
[19]于君娜.基于神经网络的高光谱遥感图像分类研究[D].哈尔滨：哈尔滨工程大学,2007.
[20]周凌翱.改进BP神经网络在模式识别中的应用及研究[D].南京：南京理工大 学,2010.
[21]贾晓东.遗传算法和BP网络相结合的分类器在BCI中的应用[D].河北：河北工业大学,2008.
[22]曹广真.多源遥感数据融合方法与应用研究[D].上海：复旦大学,2006.
[23]贾永红.人工神经网络在多源遥感影像分类中的应用[J].测绘通报,2000,(7)：7-8.
[24]Wen-jie Wang, Zhong-ming Zhao, Hai-qing Zhu. Object-oriented Multi-feature Fusion Change Detection Method for High Resolution Remote Sensing Image,2009 Urban Remote Sensing Joint Event. IEEE,2009.
[25]王惠明,史萍.纹理特征的提取方法[J].中国传媒大学学报自然科学版,2006,13(1)：49-51.
[26]侯海苗,冀小平.基于灰度共生矩阵的纹理特征[J].长治学院学报,2008,25(5)：31-31.
[27]Martin T. Hagan, Howard B.Demuth,Mark H.Beale.神经网络设计[M].北京：机械工业出版社,2006.
[28]阎平凡,黄瑞旭.人工神经网络模型分析与应用[M].合肥：安徽教育出版社,1993.
[29]王英健,戎丽霞.基于遗传BP算法的神经网络以其在模式识别中的应用[J].长沙交通大学学报,2005,(1)：54-56.
[30]杨南达,李世平.遗传算法研究[J].兵工自动化,2008,27(9)：60-62.
[31]张磊,尹文龙,张额.基于遗传算法的BP网络优化设计方法[J].科学技术与工程,2008,8(14)：3770-3772.
[32]可华明,陈朝镇等.基于Matlab的GA-BPNN遥感图像分类[J].西南科技大学学报自然科学版,2010,25(3)：55-59.
[33]Jin YQ,Wang Y. A genetic algorithm to retrieve multi-parameters of land surface roughness and soil moisture[J]. International Journal of Remote Sensing,2001,22(16):3093-3099.
[34]周明,孙树栋.遗传算法原理与应用(第一版)[M].北京：国防工业出版社,1999.
[35]李仪,陈云浩,李京.基于微种群遗传算法和自适应BP算法的遥感图像分类[J].光学技术,2005,31(1)：17-23.
[36]曹广真,金亚秋.用于Landsat ETM+和ERS-2 SAR图像融合对城区地物特征分类的BP-ANN/GA混合算法[J].电子与信息学报,2006,28(7)：1153-1159.
[37]张璎璎.遥感影像可变多组分光谱混合分析方法研究[D].湖南：中南大学,2010.
[38]徐硕,王洲.基于纹理特征和神经网络的图像识别[J].中国农学通报,2007,23(9)：590-594.
[39]郭德军,宋蛰存.基于灰度共生矩阵的纹理图像分类研究[J].林业机械与术工设备,2005,33(7).
[40]SanRen H, Cheinlchang I. A generalized orthogonalalsub space projection approach to unsupervised multi-spectral image classification[J]. IEEE Transaction on Geoscience and Remote Sensing,2000,38(6):2515-2528.
[41]Ifarraguerri Agustin,Chein Ichang. Unsupervised hyperspectral image analysis with projection on pursit[J]. IEEE Transaction on Geoscience and Reomote Sensing, 2000,38(6):2529-2538.
[42]Kim,K.S.,ZHAO Y., Large-scale Pattern Growth of Graphene Films for Stretchable Transparent Electrodes[J].Nature,2007,4(57):2758-2763.
[43]董长红Mat lab神经网络与应用[M].北京：国防工业出版社,2005.
[44]阎平凡,张长水.人工神经网络与模拟进化计算[M].北京：清华大学出版社,2005.
[2]倪金生,蒋一军,张富民.遥感图像处理理论与实践[M].北京：电子工业出版社,2008.
[3]朱述龙,朱宝山,王红卫.遥感图像处理与应用[M].北京：科学出版社,2002.
[4]王爱玲,叶明生,邓秋香MATLAB R2007图像处理技术与应用[M].电子工业出版社,2008.
[5]李爽,丁圣彦,许叔明.遥感影像分类方法比较研究fJ].河南大学学报,2002(32)：70-73.
[6]孙家炳.遥感原理与应用[M].第1版.武汉：武汉大学出版社,2003.
[7]魏海坤.神经网络结构设计的理论与方法[M].国防工业出版社,2005.
[8]王翔.基于神经BP神经网络的遥感影像模式识别方法研究[D].太原：太远科技大学,2009.
[9]甘井中.基于BP神经网络图像识别的研究[D].大连：大连理工大学,2008.
[10]李登科,张京红,戴进,遥感图像处理系统ERDAS IMAGINE及其应用[J].陕西气象,2002,(2)：20-22.
[11]刘宣江.基于神经网络的遥感影像识别[D].北京：北京邮电大学,2007.
[12]刘安斐,李弼程,张先飞.基于数据融合的多特征遥感图像分类[J].数据采集与处理,2006,21(4)：463-467.
[13]安斌,陈书海等.纹理特征在多光谱图像分类中的应用[J].激光与红外,2002,32(3)：188-190.
[14]孙艳霞.纹理分析在遥感图像识别中的应用[D].新疆：新疆大学,2005.
[15]徐青张艳等.遥感影像融合与分辨率增强技术[M].北京：科学出版社,2007
[16]王知鸷.基于基于纹理及光谱信息融合的遥感图像分类方法研究[D].西安：西安科技大学,2010.
[17]杨听,汤国安等.ERDAS遥感数字图像处理实验教材[M].北京：科学出版社,2009.
[18]张海涛.光谱与纹理特征融合的遥感图像分类方法[D].西安：西安电子科技大学,2010.
[19]于君娜.基于神经网络的高光谱遥感图像分类研究[D].哈尔滨：哈尔滨工程大学,2007.
[20]周凌翱.改进BP神经网络在模式识别中的应用及研究[D].南京：南京理工大 学,2010.
[21]贾晓东.遗传算法和BP网络相结合的分类器在BCI中的应用[D].河北：河北工业大学,2008.
[22]曹广真.多源遥感数据融合方法与应用研究[D].上海：复旦大学,2006.
[23]贾永红.人工神经网络在多源遥感影像分类中的应用[J].测绘通报,2000,(7)：7-8.
[24]Wen-jie Wang, Zhong-ming Zhao, Hai-qing Zhu. Object-oriented Multi-feature Fusion Change Detection Method for High Resolution Remote Sensing Image,2009 Urban Remote Sensing Joint Event. IEEE,2009.
[25]王惠明,史萍.纹理特征的提取方法[J].中国传媒大学学报自然科学版,2006,13(1)：49-51.
[26]侯海苗,冀小平.基于灰度共生矩阵的纹理特征[J].长治学院学报,2008,25(5)：31-31.
[27]Martin T. Hagan, Howard B.Demuth,Mark H.Beale.神经网络设计[M].北京：机械工业出版社,2006.
[28]阎平凡,黄瑞旭.人工神经网络模型分析与应用[M].合肥：安徽教育出版社,1993.
[29]王英健,戎丽霞.基于遗传BP算法的神经网络以其在模式识别中的应用[J].长沙交通大学学报,2005,(1)：54-56.
[30]杨南达,李世平.遗传算法研究[J].兵工自动化,2008,27(9)：60-62.
[31]张磊,尹文龙,张额.基于遗传算法的BP网络优化设计方法[J].科学技术与工程,2008,8(14)：3770-3772.
[32]可华明,陈朝镇等.基于Matlab的GA-BPNN遥感图像分类[J].西南科技大学学报自然科学版,2010,25(3)：55-59.
[33]Jin YQ,Wang Y. A genetic algorithm to retrieve multi-parameters of land surface roughness and soil moisture[J]. International Journal of Remote Sensing,2001,22(16):3093-3099.
[34]周明,孙树栋.遗传算法原理与应用(第一版)[M].北京：国防工业出版社,1999.
[35]李仪,陈云浩,李京.基于微种群遗传算法和自适应BP算法的遥感图像分类[J].光学技术,2005,31(1)：17-23.
[36]曹广真,金亚秋.用于Landsat ETM+和ERS-2 SAR图像融合对城区地物特征分类的BP-ANN/GA混合算法[J].电子与信息学报,2006,28(7)：1153-1159.
[37]张璎璎.遥感影像可变多组分光谱混合分析方法研究[D].湖南：中南大学,2010.
[38]徐硕,王洲.基于纹理特征和神经网络的图像识别[J].中国农学通报,2007,23(9)：590-594.
[39]郭德军,宋蛰存.基于灰度共生矩阵的纹理图像分类研究[J].林业机械与术工设备,2005,33(7).
[40]SanRen H, Cheinlchang I. A generalized orthogonalalsub space projection approach to unsupervised multi-spectral image classification[J]. IEEE Transaction on Geoscience and Remote Sensing,2000,38(6):2515-2528.
[41]Ifarraguerri Agustin,Chein Ichang. Unsupervised hyperspectral image analysis with projection on pursit[J]. IEEE Transaction on Geoscience and Reomote Sensing, 2000,38(6):2529-2538.
[42]Kim,K.S.,ZHAO Y., Large-scale Pattern Growth of Graphene Films for Stretchable Transparent Electrodes[J].Nature,2007,4(57):2758-2763.
[43]董长红Mat lab神经网络与应用[M].北京：国防工业出版社,2005.
[44]阎平凡,张长水.人工神经网络与模拟进化计算[M].北京：清华大学出版社,2005.