综合多特征的极化SAR灾害损毁建筑提取方法研究

Accurate building damage information can provide a reliable basis for disaster assessment. Taking full advantage of the various features of polarimetric SAR images can describe the objects in disaster area more comprehensively and extract damaged building more accurately. Unlike the traditional information extraction methods based on pixels which difficultly integrate various features and are easily affected by multiplicative noise, the object-oriented method this paper will be exploit to study damaged building extraction methods using polarimetric SAR images, which can integrate the polarization, statistical distribution and geometry shape features. The specific research works can be organized as follows. Firstly, in order to take into account the building orientation angle, the target reflection symmetric and the characteristics of urban vegetation, a new polarimetric target model-based decomposition method will be established to extract the polarimetric features describing targets accurately in disaster areas through integrating selective de-orientation, helix scattering suppression and the adaptive volume scattering model. Secondly, we will construct the description of polarimetric heterogeneity, statistic distribution heterogeneity and synthesized multi-features heterogeneity respectively based on the fractal network evolution algorithm. Then, a multi-feature synthesized strategy will be formulated under the condition of the statistic model parameter estimation. After that, polarimetrc SAR image segmentation model integrating multi-feature will be built to construct analytical objects. Finally, the random forests method will be adopted to classify objects and to evaluate feature importance for the damaged building recognition. In summary, the outcomes of the project will be to improve the theory and methods of polarimetric SAR model decomposition and information extraction, and to enhance the analysis capabilities of SAR disaster survey and assessment.

充分利用极化SAR图像的多种特征，能更全面地描述灾区地物并准确提取损毁建筑，为灾情评估提供可靠依据。传统提取方法基于像素分析，难以全面综合多方面特征且易受噪声影响，提取精度受限。本项目采用对象分析思想，开展综合极化、统计分布和几何形状等多特征的SAR灾害损毁建筑提取方法研究。具体研究：1）顾及建筑取向角、地物反射对称性和城区植被特点，建立综合选择性去取向、螺旋体散射抑制和自适应体散射模型的极化目标模型分解方法，准确提取灾区目标的极化特征；2）基于分形网络演化算法，分别构建极化、统计分布及多特征综合的异质度描述，根据统计分布模型的参数估计等条件制定特征综合策略，建立综合多特征的极化SAR图像分割模型，构建分析对象；3）引入随机森林，研究基于对象的损毁建筑多特征综合提取及特征的重要性评价方法。项目的开展将完善极化SAR模型分解和信息提取的理论与方法，提升SAR灾害调查评估的分析能力。

充分利用极化SAR图像的极化、统计分布和几何形状等多种特征，可以更全面地描述灾区地物并准确提取损毁建筑，为灾情评估提供可靠依据。根据“目标分解——图像分割——损毁建筑识别”这一主线，本项目开展了综合利用多特征的SAR灾害损毁建筑提取方法研究。主要研究内容和成果包括以下几个方面：1）针对城区复杂应用场景下基于模型目标分解存在的问题，顾及建筑取向角、地物反射对称性和城区植被特点，建立了综合选择性去取向、螺旋体散射抑制和自适应体散射模型的极化目标模型分解方法。2）基于分形网络演化算法，分别构建了极化、统计分布及多特征综合的异质度描述，根据统计分布模型的参数估计等条件制定特征利用策略，建立了综合多特征的极化SAR图像分割模型。3）在上述基础上，重点研究了利用改进的模型分解方法、统计特征、纹理特征的灾后极化SAR图像倒塌建筑物提取和评估。项目的研究成果发展了极化 SAR 模型分解、极化SAR图像分割分类的理论与方法，进一步提升了SAR灾害调查评估的分析应用能力。