【作者】 王知鸷；

【导师】 那彦；

【作者基本信息】 西安电子科技大学 ， 电路与系统， 2010， 硕士

【摘要】 本文研究了一种基于光谱特征及纹理特征融合的遥感图像分类方法。在本文设计的分类方法中,首先提出了一种新的适用于遥感目标分类的目标特征——幅度波谱向量,并把该特征和纹理特征进行融合,对特征分量进行逐点自适应加权,然后设计了一种性能良好的具有降维功能的隶属函数对目标的真实特征值进行投影变换,使其成为模糊特征值。然后,选定分类规则对目标进行分类,最后,对分类结果采用两种方法进行评价。对遥感图像进行实验分析,首先用本文提出的方法对不同复杂程度的仿真图像进行分类实验,并和基于不同特征的分类方法进行比较。其次是对融合前后的实际图像进行实验,比较其分类结果。最后对采用9种不同融合方法得到的图像用本文提出的方法进行分类,比较融合方法的好坏。通过实验分析,采用本文提出的方法,即有加权系数的特征融合方法,进行仿真实验,用全局评价的方法进行评估,对三类的图像,总体分类精度为92.60%。而采用未改进的方法,即没有加权系数的特征融合方法,其总体分类精度为78.48%；对五类的图像,用本文提出的方法,即有加权系数的特征融合方法,进行仿真实验,用全局评价的方法进行评估,总体分类精度为91.12%,而采用未改进的方法,即没有加权系数的特征融合方法,其总体分类精度为90.67%。对采用9种不同融合方法得到的图像进行分类,其分类结果明显好于融合前的图像。对于三类图像,融合前图像的分类精度为90.77%,而采用9种融合方法得到的图像的平均精度为92.62%；对于五类图像,融合前图像的分类精度为88.99%,而采用9种融合方法得到的图像的平均精度为90.82%。实验结果表明：新的分类方法能够获得比当前的典型处理方法性能更好的识别效果,同时融合后的图像比融合前的图像分类效果更好。更多还原

【Abstract】 A new remote sensing object classification algorithm based on the fusion of texture feature and spectral feature is proposed.In the classification algorithm which is prompted in this paper, a new kind of target feature named feature spectrum vector is put forward here. Meanwhile, a new kind of feature fusion method that based on vector relativity is prompted here. And then, a new membership function with better performance and the decreasing dimension function is promoted here. Using this function to make projection transformation to the real feature, then the fuzzy feature value can be received. And then, choose a kind of classify rule, and label the targets. At last, the classification results are evaluated by two different methods.Remote sensing images are experimented using the promoted classification algorithm proposed in this paper. First, simulation images of different levels of complexity are experimented and compared with different classification methods based on different features. Second, original image and fusion image are experimented and compare the classification results. Last, use the promoted algorithm to assort experiment of remote sensing images with nine different fusion methods and compare the classification results.Through the experimental analysis, use the promoted algorithm, ie feature fusion with the weighting coefficients, to assort experiment of real remote sensing images. For three kinds of iamge, the overall classification accuracy is 92.60%. But using the unimproved algorithm, ie feature fusion without the weighting coefficients, the overall classification accuracy is 90.15%. For five kinds of iamge, use the promoted algorithm, ie feature fusion with the weighting coefficients, to assort experiment of real remote sensing images, the overall classification accuracy is 91.11%. But using the unimproved algorithm, ie feature fusion without the weighting coefficients, the overall classification accuracy is 87.58%.The images on the use of 9 different fusion methods are classified that the classification results significantly better than before the fusion images. For three kinds of no fusion iamge, the overall classification accuracy is 90.77%, but for the fusion image, the overall classification accuracy is 92.62%; For five kinds of no fusion iamge, the overall classification accuracy is 88.99%, but for the fusion image, the overall classification accuracy is 90.82%.The result suggests that the promoted algorithm can achieve better classification performance. Meanwhile, the classification results of the fusion images significantly better than before the fusion images.更多还原