面向对象的滑坡编目快速绘制方法研究

A complete multi-temporal landslide inventory is essential for quantitative landslide hazard and risk assessment. Traditional mapping method, which relies on manual interpretation of aerial photographs and intensive field surveys, is time and labor consuming. Pixel-based analysis of satellite images, which provides a potential approach of automated inventory mapping, is nevertheless only based on spectral behaviors of landslides. With a consideration of spectral, morphological and contextual features, object-oriented analysis allows a cognitive approach supported by expert knowledge. However, traditional object-oriented methods have several limitations: (1) inaccurate segmentation of landslide objects; (2) highly reliant on manual thresholding of landslide features and (3) not generic when adapting to different locations and datasets. ..The purpose of this study is to develop a novel object-oriented approach for a rapid mapping of landslide inventory using machine learning algorithms, so as to overcome those limitations as mentioned above. The usefulness of this approach is demonstrated by a case study in HongKou County that was strongly affected by 2008 Wenchuan earthquake. The aim is to rapidly prepare an inventory of landslides triggered after the earthquake through an object-oriented analysis on VHR satellite images. The applicants hope that this approach could provide a fundamental support for future landslide inventory mapping over large areas of China.

建立完整的历史滑坡编目对开展灾害调查、总结灾害分布规律，构建完整灾害风险评估体系有重要作用。传统滑坡编目绘制方法主要基于航空照片的判读和大量野外调查，需耗费大量时间和人力。基于像素的卫星影像自动分析方法虽然快速，却只能基于滑坡的光谱特征进行编目绘制。面向对象的影像分析能完整地描述滑坡的光谱、语义、形状等特征，并能模仿人类感知识别过程进行编目绘制，但现存方法仍有以下局限性：（1）滑坡对象分割不准确；（2）滑坡特征阈值需通过反复试错来确定，具有主观性；（3）绘制方法通用性低，同一规则难以应用于不同地区或数据。本课题拟开发一种全新的面向对象的机器学习推理算法对滑坡编目进行快速绘制，以解决现存方法的上述局限性。本研究将选取受汶川地震影响的四川省虹口乡震区作为示范区域，通过对高分辨率遥感影像的分析，快速绘制震后虹口乡地区的滑坡编目。希望通过此研究，为今后我国绘制大范围完整的历史滑坡编目提供技术支撑。

建立完整的滑坡编目对开展灾害调查、总结灾害分布规律、构建完整的滑坡灾害风险评估体系至关重要。传统滑坡编目绘制方法主要基于航空照片的判读和大量野外调查，却需耗费大量时间和人力。基于像素的遥感影像自动分析方法虽然快速，却只能基于滑坡的光谱特征进行编目绘制。面向对象的影像分析能完整地描述滑坡的光谱、语义、形状等特征，并能模仿人类感知识别过程进行编目绘制，但现存方法仍有以下局限性:（1）滑坡对象分割不准确；（2）滑坡特征阈值需通过反复试错来确定，具有主观性；（3）绘制方法通用性较低，同一算法难以应用于不同地区或数据。本项目开发一种全新的面向对象的滑坡编目快速绘制方法，以解决现存方法的上述局限性。首先，本项目在分形网络演化算法的基础上开发出一种适合高分辨率遥感影像的多尺度优化影像分割方法，可以提取出适合符合滑坡特征描述的完整滑坡对象。此外，采用Level Set Evolution（LSE）算法综合光谱、几何形态、纹理对不同类型滑坡对象进行自动判别提取和分类，以此技术为基础制作了四川省阿坝州理县和香港特别行政区大屿山岛的滑坡编目。希望通过此研究,为今后我国绘制大范围完整的历史滑坡编目提供技术支撑。成果方面，项目组共发表SCI论文9篇，申请专利3项，培养研究生4名。一项主要研究成果发表在遥感领域影响因子排名第一的《Remote Sensing of Environment》（IF＝6.4）上。