联合多源震后高分遥感影像的建筑物震害特征建模与识别

Disaster monitoring of earthquake based on the post-event high resolution remote sensing image allows the bypass of various restrictions due to the dependency on the pre-event data and it is of great importance for the emergency response after disaster. However, for the special surface features of buildings, using merely optical or SAR image is not adequate for the effective and comprehensive description of the various damage features of the buildings, so it can only extract the damage buildings while cannot identify the building damage types which is the key disaster information. Therefore, by taking the advantages of multi-source data combination as well as exploring the complimentary information of post-event high resolution optical and SAR images and combining the low-level vision feature and high-level semantics together, we are intended to carry out a research on the feature modeling of building seismic damages and establish an effective way to identify building damage to break the "bottle neck" in the this field. The main research contents include: the hierarchical analysis of post-event complex scene and building detection in multi-source data with various morphological attribute profiles; the establishment of building contour projection from optical to SAR image and unified building set from multi-source data; clustering rules of double bounce with different collapse types and semantic features extraction of building collapses; the identification of building damage types based on active learning and support vector machines etc. Research results will actively promote the progression of remote sensing application technology and improve the abilities in disaster monitoring and emergency response of China.

基于震后高分辨率遥感影像进行灾情监测，可避免对震前数据依赖而带来的诸多限制，对灾后应急响应具有重要意义。然而，面对建筑物这类特殊地物，单纯依据高分光学或SAR影像，不能全面有效地描述建筑物多样的震害特征，仅能实现对震害建筑物的提取，无法识别建筑物的震害类型这一关键灾情信息。为此，本课题拟利用多源数据联合优势，通过挖掘震后高分光学与SAR影像的互补信息，结合底层视觉特征与高层语义，以建筑物的震害特征建模为突破口，建立有效的建筑物震害识别方法。研究内容主要包括：联合多种形态学属性剖面的多源数据震后复杂场景层次化解析与建筑物检测；高分光学－SAR建筑物轮廓投影与统一多源数据建筑物基元集合构建；双回波坍塌词汇聚类与建筑物坍塌语义特征提取；基于主动学习SVMs的建筑物震害类型识别等。研究成果将积极推动遥感应用技术进步，提高我国灾害监测及应急响应能力。

基于震后高分辨率遥感影像进行灾情监测，可避免对震前数据依赖而带来的诸多限制。为此，本课题通过挖掘震后高分光学与SAR影像的互补信息，结合底层视觉特征与高层语义，开展建筑物震害特征建模及识别研究。首先，为减小分割结果与实际地理对象间的差异，构建了一种新的对象置信度指标，有助于准确提取建筑物、道路等地理对象的完整轮廓；其次，通过所提出的基于形态学属性剖面的场景抽象化表示方法，实现了建筑物的有效识别；在此基础上，利用所提出的多尺度配准方法，构建了统一的光学-SAR建筑物集合；最后，结合双回拨提取坍塌语义特征，作为主动学习SVM分类器的输入特征空间。通过本项目研究，突破了基于震后高分遥感影像的建筑物震害类型识别技术瓶颈，有助于提高我国灾害监测及应急响应能力。