土地覆盖更新中训练样本的动态选取方法

Dynamic selection method of training samples in land cover renewal

针对土地覆盖数据更新的任务,该文提出了训练样本的动态获取的方法。基于变化检测的原理和相对保守的分割阈值获取到的特定数量不变像元,在继承了历史土地覆盖标签后可以作为目标影像分类的训练样本。然而,分割阈值的确定具有较大随意性。该文的方法逐步增加不变像元的规模(即每种地类中不变像元所占地类像元总数的百分比),同时不断执行目标影像监督分类;以前后两次监督分类结果一致性水平作为不变像元规模是否合理的依据,当一致性水平达到设定水平时,提取过程结束。实验结果表明,该方法克服了训练样本选取时阈值设定的主观性,同时可以避免训练样本中地类缺失的可能性。

For the task of updating land cover data,this paper proposed a method for dynamic acquisition of training samples.The specific number of invariant pixels obtained based on the principle of change detection and the relatively conservative segmentation threshold can be used as training samples for target image classification after inheriting the historical land cover label.However,the determination of the segmentation threshold has greater arbitrariness.The method of this paper gradually increased the size of the invariant pixels(that is,the percentage of the total number of pixels in the invariant pixels in each type of land),and continuously performed the target image supervised classification.The level was used as the basis for the reasonable size of the invariant pixels.When the consistency level reached the set level,the extraction process ended.The experimental results showed that the method overcame the subjectivity of the threshold setting when training samples were selected,and could avoid the possibility of missing the ground in the training samples.