Abstract: In order to make full use of spectral information and spatial structure information of hyperspectral image, a new classification approach based on random forest for hyperspectral data classification is proposed. Firstly, the dimension of the data is reduced by principal component analysis. The spectral features are extracted by independent component analysis, which can eliminate the spatial correlation among pixels. Then morphological analysis is used to extract the spatial structure features of pixels. Moreover, random forests are respectively constructed based on spectral information and spatial structure feature, and the spatial continuity constraint is introduced to fix the prediction results of pixels. Finally, the classification result is decided by voting strategy. Experiments on AVIRIS and ROSIS hyperspectral images demonstrate that the proposed approach has better performance than the classical methods and methods based on single feature.

Key words: hyperspectral image, image classification, random forest, independent component analysis

摘要： 为了充分利用高光谱图像的光谱信息和空间结构信息，提出了一种新的基于随机森林的高光谱遥感图像分类方法，首先，利用主成分分析降低数据的维数，并对主成分进行独立成分分析提取其光谱特征，同时消除像元的空间相关性，再采用形态学分析提取像元的空间结构特征，然后，根据像元的谱域和空域特征分别构造随机森林，并引入空间连续性对像元点的预测结果进行约束修正，最后由投票机制决定最后的分类结果。在AVIRIS和ROSIS高光谱图像上的实验结果表明，所提方法的分类性能要优于传统的高光谱图像分类方法，且分类精度高于基于单一特征的方法。

关键词: 高光谱图像, 图像分类, 随机森林, 独立成分分析