多尺度超像素纹理特征保持与融合的高光谱图像分类

Hyperspectral Image Classification Based on Multi-scale Superpixel Texture Preservation and Fusion

高光谱图像的低空间分辨率特性往往导致全局纹理提取技术难以获取地物要素的精准纹理信息,同时,单一尺度的局部纹理提取技术难以达到有效识别地物的目的。基于此,该文设计了一种多尺度超像素纹理保持与融合(MSuTPF)的高光谱图像分类方法,主要架构如下:首先,利用2D Gabor滤波器对高光谱图像进行多方向与尺度的全局纹理提取,并通过融合各尺度的纹理特征,增强纹理结构表征能力;其次,融合纹理与光谱主成分特征以形成光谱-纹理联合判别特征;再次,采用形状自适应的超分割方法,作用至光谱-纹理联合特征进行局部纹理信息保持与融合,尤其是,为克服超像素邻域像元的隐性不相关问题,该文定义了基于密度最近邻相似性评价准则,使超像素纹理进一步趋于一致性;最后,将各更新的光谱-纹理联合特征输入像素级分类器获取其对应的类标签,并采用多数表决的决策融合机制取得最终分类结果。Indian Pines和Pavia University真实数据集的实验表明,该方法在小样本条件下的分类精度优于基准分类器(SVM)、深度学习方法(GFDN)以及最新的空-谱分类方法(S3-PCA)等8个对比方法,充分证明了该文所提方法的实用性和有效性。

The low spatial resolution characteristics of hyperspectral images often make it difficult for global texture extraction techniques to obtain accurate texture information and the single-scale local texture extraction technology is not satisfactory for effectively identifying the features.In this article,a Multi-scale Superpixel Texture Preservation and Fusion is proposed for hyperspectral image classification.Specifically,the original hyperspectral image is first extracted with multi-direction and scale global texture using 2D Gabor filter,and the texture feature of each scale is merged to enhance the texture structure characterization ability.Next,texture and spectral principal component features are fused to form spectral-texture joint discriminant features.After that,the shape adaptive oversegmentation method is applied to the spectral-texture joint feature for local texture information preservation and fusion.In particular,in order to overcome the hidden irrelevance problem of neighboring pixels,a density-based nearest neighbor similarity evaluation criterion is defined,which aims to make the superpixel texture more consistent.Finally,the updated spectral-texture joint discriminant features are input into the pixel-level classifiers to obtain their corresponding class labels,and the decision fusion mechanism of majority voting is adopted to obtain the final classification result.Experiments on the real data sets of Indian Pines and Pavia University show that the classification accuracy of this method under the condition of small samples is better than eight comparison methods such as the benchmark classifier Support Vector Machine(SVM),deep learning method Gabor Filtering and Deep Network(GFDN),and the latest spatial-spectral method Spectral-Spatial and Superpixelwise Principal Component Analysis(S3-PCA),which proves fully the practicability and effectiveness of the proposed method.