基于二维奇异谱特征提取的高光谱影像同质划分

Homogeneous Area of the Hyperspectral Image Based on Two-Dimensional Singular Spectrum Feature Extraction

高光谱影像数据具有维度高、信息冗余等特征,传统的特征提取方法通常使用了固定窗格提取高光谱影像的空间特征,忽略了地物之间的空间关系,对地物空间信息利用不充分。对此,文章提出了融合超像素算法的二维奇异谱分析方法,通过超像素划分并提取同质区域,经过二维奇异谱分析,从每个波段中提取空间结构信息,增强本类别的特征信息,同时减少类间差异性与噪声带来的影响。对所提取的空间特征,结合粒子优化算法提取高光谱影像最佳波段组合。实验结果表明,在Indian Pines与Salinas数据影像数据集中,使用同样的支持向量机分类器,文章所提出特征提取方法实现的分类精度相比于原始数据分别提升了15.99%与3.7%,相比于2DSSA提升了3.12%与0.91%。改进的奇异谱分析方法,可以充分利用同质区域的局部一致性,从而提高数据处理的性能,减少了影像中的冗余信息和噪声。

Hyperspectral image data has the characteristics of high dimension,information redundancy,etc.Traditional feature extraction methods usually use fixed pane to extract the spatial features of hyperspectral image,ignore the spatial relationship between ground objects,and make insufficient use of the spatial information of ground objects.In this paper,a two-dimensional singular spectrum analysis method based on superpixel algorithm is proposed.The homogeneous region is divided and extracted by superpixel.After two-dimensional singular spectrum analysis,spatial structure information is extracted from each band to enhance the characteristic information of this category and reduce the influence of inter-class differences and noise.The optimal band combination of hyperspectral image is extracted by particle optimization algorithm based on the extracted spatial features.The experimental results show that in Indian Pines and Salinas data image sets,the classification accuracy achieved by the proposed feature extraction method is improved by 15.99%and 3.7%compared with the original data,and by 3.12%and 0.91%compared with 2DSSA,respectively,using the same SVM classifier.The improved singular spectrum analysis method can make full use of the local consistency of the homogeneous region,thus improving the performance of data processing and reducing the redundant information and noise in the image.