高分辨率遥感影像分类的多示例集成学习

Multiple instance ensemble learning method for high-resolution remote sensing image classification

精选示例特征嵌入多示例学习(MILES)算法在对噪声较强的训练样本进行学习时表现出良好的性能,但其判断规则可能带来遥感影像分类结果的不确定性。针对这一问题,提出用Bagging和AdaBoost集成MILES的多示例集成学习算法,使用粗包细分、多样性密度和最大似然分类相结合抑制分类不确定性的方法,实现了高分辨率遥感影像分类中多示例学习与集成学习的组合。采用Quick Bird、IKONOS等高分辨率遥感影像进行试验,结果表明多示例集成学习能有效控制遥感影像分类结果的不确定性,具有良好的应用前景。

Multiple Instance Learning Via Embedded Instance Selection （MILES） has shown good performance in dealing with noisy training samples, but its bag prediction rule may introduce new uncertainty into the remote sensing image classification results. In order to overcome this limitation, two popular ensemble learning strategies, Bagging and AdaBoost are integrated with MILES. Two methods are proposed to constrain the uncertainty in remote sensing image classification： re-classification of coarse bags, and integration of MILES with diverse density and maximum likelihood classifier. The experimental results show that the uncertainty of remote sensing image classification can be obviously reduced by the integration of multiple instance learning with ensemble learning.