高光谱图像分类的ReliefF-RFE特征选择算法构建与应用

高光谱图像具有波段连续、维数高、数据量大、相邻波段相关性强的特点，可为地物分类提供更为丰富的细节信息。但是，数据中存在大量冗余信息与噪声，在图像分类中如直接利用其所有波段特征而不进行有效分析与选择，将会导致较低的计算效率和较高的计算复杂度，分类精度亦可能随着波段维数增加而出现先增后减的“休斯（Hughes）现象”。为快速地从高达数十个甚至数百个波段的高光谱图像中提取出具有较好识别能力的特征子集，从而避免“维度灾难”，将过滤式ReliefF算法和封装式特征递归消除算法（RFE）相结合，构建了ReliefF-RFE特征选择算法，可用于高光谱图像分类的特征选择。该算法根据权重阈值，利用ReliefF算法快速剔除大量无关特征，缩小并优化特征子集的范围；利用RFE算法进一步搜索最优特征子集，将缩小范围后的特征子集中与分类器关联性小、冗余的特征进行递归筛选，进而得到分类性能最佳的特征子集。采用Indian pines数据集、Salinas-A数据集与KSC数据集等3个标准数据集作为实验数据，将ReliefF-RFE算法的应用效果与ReliefF和RFE算法进行对比。结果显示，在3个数据集中，应用ReliefF-RFE算法的高光谱图像分类平均总体精度（OA）为92.94%、F-measure为92.81%，Kappa系数为91.94%；ReliefF-RFE算法的平均特征维数是ReliefF算法的37%，而平均运算时间则是RFE算法的75%。由此表明，ReliefF-RFE算法能够在保证分类精度的同时，克服过滤式ReliefF算法无法有效减小特征之间冗余以及封装式RFE算法时间复杂度较高的缺陷，具有更为均衡的综合性能，适用于高光谱图像分类的特征选择。

Construction and Application of ReliefF-RFE Feature Selection Algorithm for Hyperspectral Image Classification

Hyperspectral images are characterized by continuous bands, high dimensionality, large data volume and strong correlation between adjacent bands, which can provide richer detailed information for feature classification. However, there is a lot of redundant information and noise in data, and the direct use of all band features without effective analysis and selection in image classification will lead to low computational efficiency and high computational complexity, and the “Hughes phenomenon” that the classification accuracy may increase and then decrease with the increase of band dimension. In order to quickly extract a subset of features with good recognition ability from hyperspectral images with tens or even hundreds of bands to avoid the “dimensional disaster”. This paper combines the filtered ReliefF algorithm and the wrapped recursive feature elimination algorithm (Recursive feature elimination, RFE) to build the ReliefF-RFE feature selection algorithm, which can be used for feature selection in hyperspectral image classification. The algorithm uses the ReliefF algorithm to quickly eliminate many irrelevant features based on weight thresholds to narrow and optimize the range of feature subsets. The RFE algorithm is used to further search for the optimal feature subsets, and the recursive elimination of the less relevant features and redundant to the classifier in the narrowed feature subsets is performed to obtain the feature subsets with the best classification performance. In this paper, three standard datasets, including the Indian pines dataset, Salinas-A dataset and KSC dataset, are used as experimental data to compare the application effect of the ReliefF-RFE algorithm with ReliefF and RFE algorithms. The results show that the hyperspectral image classification by applying the ReliefF-RFE algorithm has an average overall accuracy (OA) of 92.94%, F-measure of 92.81%, and Kappa coefficient of 91.94%; in the three datasets, the average feature dimension of ReliefF-RFE algorithm is 37% of that of ReliefF algorithm, while the average operation time is 75% of that of the RFE algorithm. It shows that the ReliefF-RFE algorithm can ensure the classification accuracy while overcoming the defects of the filtered ReliefF algorithm, which cannot effectively reduce the redundancy among features and the wrapped RFE algorithm, which has high time complexity and has a more balanced comprehensive performance, which is suitable for feature selection in hyperspectral image classification.