单类分类器应用于多源遥感数据的关键问题研究：以农田提取为例

Remotely sensed data have been widely used for land use/cover classification and thematic mapping. However, traditional supervised classification requires correct training sample for all the classes, which is very challenging in practice. On the other hand, many applications focus only on one specific target class. Therefore, one-class classifiers received more and more attentions in remote sensing community recently, because they require only the training sample of the target class, which is much more convenient in applications. PUL (Positive and Unlabeled Learning) is a typical one-class classifier and has been validated in classification for single-scene remotely sensed imagery. However, there are still two challenges for applying PUL to multi-source remotely sensed data: 1) PUL assumes that the training sample should be selected randomly from target class; unfortunately, this assumption is difficult to be satisfied in reality; 2) Because of the cloud contamination and hardware limitation, remotely sensed data from different sources usually cover different areas, which leads to feature missing for most pixels; however, the issue of feature missing cannot be addressed by available PUL algorithm. This project aims to address these two problems by a case study of cropland mapping. A novel PUL algorithm, which can be used for nonrandom training sample and multi-source remotely sensed data with feature missing issue, is expected to be developed. Finally, the proposed PUL will be used for mapping cropland in China.

在遥感影像解译中，传统的监督分类方法需要正确选取所有类别的训练样本，这在实际操作中往往很难做到。另一方面，很多应用或专题制图只对特定的目标类别感兴趣。而单类分类器只需要目标类别的训练样本，十分符合实际应用中的需求，近来逐渐受到遥感学者的重视。其中，PUL（Positive and Unlabeled Learning）算法是一种典型的单类分类算法，其有效性已经在单景遥感影像分类上得到验证。然而，要将PUL广泛应用于多源遥感数据，还存在两个问题：1）PUL算法假设目标类别的训练样本从目标类别中完全随机选取，而通常的训练样本数据很难满足这样的假设；2）由于云污染、传感器硬件限制等原因，多源遥感数据的有效覆盖范围往往不一致，导致大部分像元存在特征缺失的问题，而PUL尚不能直接应用于该情况。本研究拟以农田提取为例，探讨这两个问题的解决方案，并将改进的PUL应用于全国范围内的农田分布制图。

单类分类器正样本与未标定学习（Positive and Unlabeled Learning, PUL）已经在遥感影像分类上得到验证，由于其只需要特定感兴趣类别的训练样本，因此对于农田提取具有重要的应用价值。但是PUL农田制图的广泛应用还有赖于厘清其中训练样本的影响以及遥感影像云污染的问题。本项目就PUL农田制图中的几个重要问题进行探讨：1) 提出迭代式云雾最优变换（Iterative Haze Optimized Transformation，IHOT），通过多时相数据的相关信息，利用迭代回归进行精准云检测，为遥感耕地分类提供高质量的数据输入；2)　研究了训练样本的数量、质量与纯度对PUL分类器的影响，建议了训练样本的合理选取策略; 3) 针对纯样本更倾向于被选择的现实情况, 提出一种迭代式的算法来增加正样本的数量，并纠正人工选取时过的纯样本; 4) 支持了全球农田制图的部分工作。