融合多特征与互信息选择集成多核极限学习机的影像分类方法

针对影像分类结果的类间差异性与准确性难以平衡的问题,提出一种融合多特征与互信息选择集成多核极限学习机的影像分类方法。该方法首先利用最小噪声分离提取影像的光谱特征,考虑到高分辨率影像局部细节信息清晰,利用LBP算子提取影像的局部纹理信息,采用泛化性能好的核极限学习机训练多个弱分类器;然后,通过引入相关性准则描述准确性,冗余性准则描述差异性,将选择性集成多核极限学习机问题转化为变量选择问题;最后,利用基于互信息的最大相关最小冗余准则,对生成的多核极限学习机进行选择,从而实现影像分类结果差异性与准确性的平衡。文章采用高分二号数据实验,总体分类精度和Kappa系数分别为92.03%、0.9。分析结果表明,该方法能够利用多种特征的分类优势,进而有效改善了高分二号影像的分类结果。

Image Classification Method Based on Multiple Kernel Extreme Learning Machine Integrated by Multi Features and Mutual Information Selection

Aiming at the problem that the difference and accuracy of image classification results are difficult to balance,a new image classification method is proposed,which integrates multi-core learning machine with multi feature and mutual information selection.In this method,firstly,the spectral features of the image are extracted by the minimum noise separation,and the local texture information of the image is extracted by LBP Operator,considering that the local details of the high-resolution image are clear.Then,by introducing correlation criteria to describe the accuracy and redundancy criteria to describe the difference,the problem of selective integration of multiple kernel extreme learning machine is transformed into the problem of variable selection.Finally,use the maximum correlation minimum redundancy criterion based on mutual information to select the generated multiple kernel extreme learning machine,so as to achieve the balance between the difference and accuracy of image classification results.In this paper,GF-2 data are used for experiments,and the overall classification accuracy and Kappa coefficient are 92.03%and 0.9,respectively.The analysis results show that the proposed method can utilize the classification advantages of multiple features,and improve the classification results of GF-2 images effectively.