基于深度特征迁移学习的多源遥感影像动态阈值多类别变化检测

Multiple change detection from multi-source remote sensing images has become a hot topic in remote sensing information processing with the popularization of remote sensing data of multi-platform sensors. Due to the discrepancies of imaging mechanism, different spectral, spatial resolution and radiation difference are common among multi-source images, which cannot be directly compared to generate the difference image. As such, it is difficult for traditional change detection methods based on analysis of the difference image to detect multiple changes from multi-source images. In view of the above problems, this project intends to study multiple change detection using dynamic thresholds based on transfer learning of deep features from multi-source remote sensing images. The main research contents of this project include: (1) extract the direct and comparable features of multi-source images to accurately describe changes by introducing the depth feature expression and transfer learning; (2) establish a dynamic threshold model in the polar framework for multiple change detection with regard to the characteristics of each change category; (3) build a multi-phase active contour model which integrates local and global information at the same time, and refine the detection results with dynamic thresholds. On the basis of research results, an efficient multiple change detection scheme for multi-source remote sensing data including feature extraction, dynamic threshold decision and post-processing. This research will effectively solve the problem in multiple change detection from multi-source remote sensing images, thereby improve its efficiency and the reliability of results, and promote the integrated applications of multi-source remote sensing data.

随着多平台传感器遥感数据的普及应用，多源遥感影像的多类别变化检测已成为遥感信息处理领域的研究热点。由于成像机理的不同，多源影像间往往存在明显的光谱、空间分辨率及辐射差异而无法直接比较生成差异影像，导致传统基于差异影像分析的变化检测方法难以适用于多源影像的多类别变化检测。针对上述问题，本项目拟研究基于深度特征迁移学习的多源影像动态阈值多类别变化检测。包括：（1）引入深度特征表达及迁移学习理论，提取多源影像间准确描述地物变化的直接可比特征；（2）顾及各变化类别特性，在极坐标框架下构建动态阈值多类别变化检测模型；（3）研究同时集成局部和全局信息的多相主动轮廓模型，实现阈值检测结果的精化。在此基础上，形成一套包括特征提取、动态阈值设置和后处理等流程化的多源遥感影像多类别变化检测方案。本项目可有效解决多源影像多类别变化检测的难点问题，提高其整体效率和检测结果的可靠性，推动多源遥感数据的集成应用。

针对多源遥感影像受光谱、空间分辨率及辐射差异限制难以通过直接比较生成差异影像进行多类别变化检测的难题，引入深度学习和迁移学习技术，开展多源影多类别动态阈值变化检测研究。具体研究成果包括：（1）根据多源遥感影像中地物结构相似性，引入形状上下文，构建地物结构相似性描述符，实现了多源遥感影像的高精度自动配准。（2）提出一种基于深度特征迁移学习的异源遥感影像半自动监督变化检测方法，通过构建训练样本自动选择模型，提高了深度学习效率和变化检测精度。（3）利用多种特征组合，充分考虑不同变化类别的特性，提出一种动态阈值的多类别变化检测模型，并利用多相主动轮廓模型进行优化，提高了多类别变化检测的精度。（4）同时顾及全局与局部信息，构建了一种优势融合的变化检测框架，通过融合阈值和聚类变化检测算法的优势，提高了变化检测结果的精度。（5）将变化检测算法应用于遥感时空数据融合，提出一种改进灵活时空数据融合方法，增强了融合算法的可靠性和稳定性。研究成果形成了适用于多源遥感影像多类别变化检测的新方法，提高了检测检测结果的可靠性，有效推动了多源遥感影像的集成应用。