结合Beltrami流和递归滤波的高光谱图像分类方法

提出一种结合Beltrami流滤波和域转换递归滤波的高光谱图像分类算法(BFRF-SVM)。分别利用Beltrami流对主成分分析(PCA)降维后的高光谱图像滤波方法和域转换递归滤波方法对全光谱波段进行滤波,两种空间信息进行线性融合后交由支持向量机(SVM)完成分类。实验表明,相比使用光谱信息、高光谱降维、空谱结合的SVM分类方法,以及边缘保持滤波和递归滤波以及形态学滤波特征方法,本文提出的BFRF-SVM方法对高光谱图像的分类精度有较大提高,验证了该方法的有效性。

Hyperspectral Image Classification Method Combined Beltrami Flow and Recursive Filter

In the past, spatial feature extraction of hyperspectral image was usually limited to one feature extraction, and the more comprehensive spatial feature was not obtained. An improved scheme was put forward according to existent methods. An algorithm of classification (BFRF-SVM) was proposed, which was combined with spatial information obtained by Beltrami flow and domain transform recursive filter. Firstly, the spatial feature was extracted by Beltrami flow on hyperspectral data whose dimensions were reduced by principal component analysis (PCA), and the spatial correlation feature was obtained by domain transform recursive filter for all bands. Secondly, the two kinds of feature were combined, which were classified by SVM. The BFRF-SVM classification method was implemented on the hyperspectral data of Indian Pines and Pavia. The following results were obtained. In the first place, the overall accuracy (OA) of Indian Pines was 96.01% and that of Pavia was 97.46%, which were 12~15 percentage points higher than that of SVM, 12~16 percentage points higher than that of PCA-SVM, 2~12 percentage points higher than that of SGB-SVM, SBL-SVM and SGD-SVM, 4~5 percentage points higher than that of EPF, 1~3 percentage points higher than that of IFRF, and 2~6 percentage points higher than that of SMP-SVM, respectively, showing very good performance in hyperspectral classification. In the second place, although the training samples were only 7% of Indian Pines and 3% of Pavia, the OA of both can reach 96.01% and 97.46%, respectively, which removed the salt and pepper noise in the classification map obviously. In the last place, although the training samples were reduced to 4% and 0.5% for Indian Pines and Pavia, the OA can be over 91% and 90%, respectively. When the training samples were increased to 10% and 4.5%, the OA can exceed 97% and 98%, respectively. The effectiveness of BFRF-SVM was fully verified in the hyperspectral classification with good stability. The experiments showed that the BFRF-SVM algorithm was better than original SVM with the pure spectrum information, dimensionality reduction, the spatial-spectral information, the method of edge-preserving filtering and recursive filtering, and the morphological feature based method. The performance of hyperspectral image classification algorithm, i.e. BFRF-SVM, was greatly improved, and the effectiveness of the method was fully verified. The method can be applied into the field of classification and identification for agriculture and forest.