迁移学习支持下的遥感影像对象级分类样本自动选择方法

面向遥感大范围应用的目标，自动化程度仍是遥感影像分类面临的重要问题，样本的人工选择难以适应当前土地覆盖信息自动化提取的实际应用需求。为了构建一套基于先验知识的遥感影像全自动分类流程，本文将空间信息挖掘技术引入到遥感信息提取过程中，提出了一种面向遥感影像对象级分类的样本自动选择方法。该方法通过变化检测将不变地物标示在新的目标影像上，并将过去解译的地物类别知识迁移至新的影像上，建立新的特征与地物关系，从而完成历史专题数据辅助下目标影像的自动化的对象级分类。实验结果表明，在已有历史专题层的图斑知识指导下，该方法能有效地自动选择适用于新影像分类的可靠样本，获得较好的信息提取效果，提高了对象级分类的效率。

An Automatic Sample Collection Method for Object-oriented Classification of Remotely Sensed Imageries Based on Transfer Learning

For the large-scale remote sensing applications, the automatic classification of remotely sensed imageries is still a challenge. For example, the artificial sample collection scheme cannot meet the needs of automatic information extraction from the remotely sensed imageries. In order to establish a prior knowledge-based and fully automatic classification method, an automatic sample collection method for object-oriented classification, with the introduction of data mining to the process of information extraction, is proposed. Firstly, the unchanged landmarks are located. Then the prior class knowledge from old interpreted thematic images is transferred to the new target images. And the above knowledge is then used to rebuild the relationship between landmark classes and their spatial-spectral features. The results show that, with the assist of preliminary thematic data, the approach can automatically obtain reliable object samples for object-oriented classification. The accuracy of the classified land-cover types and the efficiency of object-oriented classification are both improved.