利用目标类样本的遥感分类策略

Classification strategy for remote sensing data based on positive training samples

针对传统遥感分类要求训练集涵盖所有表观地物及亚类,对样本选取要求颇高,而实际应用中并不能保证所有类别都被标记的问题,该文提出基于目标类样本的遥感循环分类策略。该方法仅输入少量目标类样本,采用正样本、未标记样本的分类算法,实现在较少样本数量和种类的条件下,对目标类稳定、高效地识别。利用10组航空影像数据,建立与传统遥感分类对比实验,结果表明:相同实验条件下,基于目标类样本的遥感分类策略与传统遥感分类具有相当的分类效果;当样本集不完全时,该策略具有更稳定的高精度识别,总体精度与Kappa系数平均分别有5.2%和7.2%的提高。该方法能够有效解决不完全训练集分类问题。

For traditional classification methods,users need to provide training sample set with all the land types existing in the image.But actually,many applications are dealing with incomplete training samples that are lack of some types.In this paper,a classification strategy based on the positive training samples was proposed.The positive and unlabeled learning algorithm was used to realize high accuracy remote sensing classification using less samples on amounts and types.To verify the correctness and advantage of the new strategy,10 independent experiments with different study sites were used to build comparative analysis.When the training sample set contained all the land-types,the accuracy assessment of the new strategy had the same statistic sense as the traditional methods;and when there are some types missing in the training set,the performance was much better than the traditional method with an average increase of5.2%at overall accuracy and 7.2%at kappa coefficient.The new strategy has the better performance when dealing with incomplete training samples.