基于对象多尺度特征深度学习的遥感影像变化检测方法研究

The research of the object-oriented remote sensing image change detection is very popular now. There are two main methods to access the object, one is based on image segmentaion, the other is based on conflation with vector data. Image segmentation is a generative information processing method based on the raw data, meanwhile, vector data conflation is a discriminative information processing method based on the top-level semantics. Integration of the two methods need to step over the semantics gap, which makes remote sensing image analysis difficult. The project combines the two methods to obtain multi-scale image object. The multi-scale image segmentation method with adaptive selection of optimal scale is used to extract multi-scale objects between the pixels and top-level object. On this basis, a multi-scale object feature deep learning method based on deep belief network is proposed to mining the ability of expression objects’ semantics with multi-scale features and building the mapping between the multi-scale features and objects’ semantics. In the end, a spatial autocorrelation method is used to examine the multi-temporal mapping consistency between the multi-scale features and objects’ semantics, and identify the change objects. Through the study of these key technologies, a novel method is provided to the application of the remote sensing change detection.

面向对象遥感影像变化检测是当前研究的热点之一。现有对象的获取方法一是基于影像分割的自底向上的获取，二是基于矢量数据套合的自顶向下的获取。分割法属于基于底层数据的产生式信息处理方法，套合法属于基于顶层语义的判别式信息处理方法，二者之间存在语义鸿沟，给遥感影像分析带来困难。本项目结合两种像斑获取方式，采用具有自适应空间尺度选择的多尺度分割方法，建立底层像元到顶层对象之间的多尺度像斑对象，使其具有丰富的多尺度信息和语义信息。在此基础上，引入深度学习方法，提出一种基于深度信念网络的多尺度像斑对象特征深度学习方法，充分挖掘对象不同尺度特征对对象语义信息的表达能力，建立对象多尺度特征与语义之间的映射关系。利用空间自相关方法对不同时相对象多尺度特征与语义映射关系的一致性进行检查，识别变化对象。通过上述关键技术的突破，为实现基于对象的遥感影像分析方法和变化检测应用提供一条新的技术途径。

深度学习是当前机器学习领域的研究热点，在语音识别、图像处理等领域具有良好的应用前景。本研究以深度卷积网络为基础，总结分析了Caffenet、Fast model、Alexnet和Fully Convolutional Networks等现有流行的深度学习网络模型，针对MNIST、CIFAR10和UC Merced Land Use等不同类型图像数据集开展了深度学习训练和图像识别研究。通过对网络不同层级的权重进行可视化，实现了影像对象多尺度特征抽取和语义一致性检验。在此基础上提出了一种基于深度学习的遥感影像变化检测方法，完成了基于对象多尺度特征深度学习的遥感影像变化检测方法研究。本研究为深度学习方法在遥感图像处理领域的应用研究提供了重要的启示性线索。项目资助发表EI论文1篇，待发表论文2篇（SCI、EI）。培养硕士研究生2名，其中2名在读。项目投入经费25万元，支出16.003万元，各项支出基本与预算相符。剩余经费8.997万元，剩余经费计划用于本项目的后续研究和成果发表。