提出一种基于多分类器协同训练的遥感图像检索方法, 该方法在不同特征集上分别建立分类器, 利用不同分类器的协同性自动标记未知样本, 从而有效解决了小样本问题。通过与相关反馈方法进行实验比较分析, 结果表明, 这两种方法各有优劣, 检索结果基本相当, 然而多分类器协同训练方法避免了相关反馈过程中人工的多次反馈, 自动化程度更高。

There are usually few training samples in the tasks of content-based remote sensing image retrieval, which will lead to over-learning problem while using this small data set for training. In this paper a novel approach using co-training in multiple classifier systems is presented, which can label the unclassified samples automatically by using the cooperative determination of the classifiers which are created on several different feature sets, so that the small sample problem can be raveled out. Compared with the technique of relevance feedback, the experiments indicate that they have their own strengths and can obtain almost the same results. However, the proposed approach of co-training in multiple classifier systems is superior in regard of avoiding the needs of human intervention through relevance feedback.