Abstract: As a classic algorithm, Transfer Component Analysis（TCA） has been applied in many different fields. However, due to the large matrix operations involved, TCA is not directly applied to the classification of remote sensing images. In this study, the traditional homogenization region information is used to improve the traditional transfer component analysis idea. A randomly selected point represents a target domain homogenous region to realize one process oftransfer component analysis and classification, and using the majority results of multiple classification as the common class of target domain homogenous region. Experiments on two sets of hyperspectral images demonstrate the effectiveness of the new method.

Key words: hyperspectral image classification, transfer learning, domain adaptation, image segmentation, homogenous region

摘要： 作为一种经典迁移学习算法，迁移成分分析（TCA）已在多种不同的领域得以应用。然而，由于涉及大的矩阵运算，TCA尚无法直接用于对遥感影像进行分类。该研究利用影像同质区信息对传统的迁移成分分析思想进行改进，以随机选取的一点代表一个目标域同质区斑块实现一次迁移成分分析及分类，用同一斑块中像素多次分类结果的众数作为目标域同质区斑块的共同类别。对两组高光谱图像的实验结果证明了该方法的有效性。

关键词: 高光谱图像分类, 迁移学习, 领域自适应, 图像分割, 同质区