面向对象的城市绿化覆盖多时相高分辨率遥感图像协同识别

Urban green space plays a key role in building an eco-city and improving the quality of living environment. Monitoring the information of urban green cover is of great significance for urban ecological research and urban planning and management. The update of urban green cover information can be achieved by using multitemporal high-resolution remote sensing image classification. However, the multitemporal image classification has ill-posed problem and dataset shift problem. The training samples of one phase cannot be applied to another phase. This proposal explores a new object-oriented co-classification method for multitemporal high-resolution remote sensing images based on the co-training theory of semi-supervised learning. First, obtaining and optimizing multitemporal objects based on multitemporal high-resolution remote sensing image segmentation. Second, according to the characterization differences of urban green cover objects on multitemporal images, the single-temporal multi-view is extended to multitemporal multi-view. Third, a multitemporal co-training is performed on unlabeled samples based on multi-view. The dataset shift problem in multi-phases is transformed into the driving force of multi-view learning. Finally, a co-classification model of multi-temporal remote sensing images is established to simultaneously solve the dataset shift and ill-posed problem of multitemporal remote sensing image classification, and to obtain multi-temporal urban green cover and its types with high precision and consistency under small sample conditions. The proposed method will provide scientific support for automatic renewal of high-resolution land cover maps.

城市绿地在建设生态城市、提高人居环境质量中起着关键作用，城市绿化信息的监测对城市生态研究、城市规划管理具有重要意义。城市绿化覆盖信息的更新可通过多时相高分辨率遥感图像分类实现，但多时相图像分类存在不适定问题和数据偏移问题，一个时相的训练样本无法应用于另一个时相。本项目利用半监督学习的协同训练理论，探讨面向对象的多时相高分辨率遥感图像协同分类新方法。首先利用多时相高分辨率遥感图像分割获取和优化多时相对象；然后根据城市绿化覆盖对象在不同时相图像上的表征差异，将单时相多视图扩展为多时相多视图；通过多视图对未标记样本进行多时相协同训练，使多时相图像分类中的数据偏移问题转化为多时相协同学习的驱动力；从而建立多时相遥感图像协同分类模型，同时解决多时相图像分类中的不适定问题和数据偏移问题，在小样本条件下获取高精度、高一致性的多时相城市绿化覆盖及其类型信息，为高分辨率土地覆盖图自动更新提供科学支撑。

城市绿化覆盖的监测对城市生态研究、城市规划管理具有重要意义，其更新可通过多时相高分辨率遥感图像分类实现，但多时相图像分类存在不适定问题和数据偏移问题，一个时相的训练样本无法应用于另一个时相。本项目基于半监督学习的协同训练理论，探讨面向对象的多时相高分辨率遥感图像协同分类新方法。主要研究内容包括利用多时相遥感图像分割获取多时相对象、基于多时相对象进行多视图表达、通过协同训练实现多时相城市绿化覆盖对象识别三大部分。项目成功建立了多时相遥感图像协同分类模型，利用不同时相数据之间的差异驱动高置信度未标记样本的选择，促进多模型之间的相互学习，提升了模型的预测能力。多时相遥感图像协同分类模型能够同时解决多时相图像分类中的不适定问题和数据偏移问题，在小样本条件下获取高精度、高一致性的多时相城市绿化覆盖及其类型信息，为高分辨率土地覆盖图自动更新提供科学支撑。