基于非规则图形标识点过程的高分辨率遥感影像几何特征提取方法研究

Remote sensing images acquired with the last generation high spatial resolution sensors provide great geometric precision and a high level of thematic detail of objects on the Earth surface. The significant amount of geometric details presented in a fine scene makes it possible to extract geometric features hidden in these high spatial resolution images. But at the same time, the improvement in spatial resolution increases the internal spectral variability of each land cover class, decreases the spectral variability between different classes, and induces geometric noise through the land cover caused by tiny targets on it. Thus, the resulting defects from high spatial resolution lead to the difficulty in the extraction of the geometric features. Accordingly, the project will address the development of novel techniques for solving the problem..Stochastic models based on marked point processes have proven to be powerful tools to deal with geometric feature extraction problems from high spatial resolution remote sensing images, and have already led to convincing experimental results in various applications such as extraction of buildings, road networks, and tree crowns. The marked point processes exploit random variables whose realizations are configurations of geometric objects, e.g., rectangles, segments, or ellipses. It has be shown that remote sensing image representations produced by these stochastic models are particularly suitable for solving geometric feature recognition problems. However, in all of the works available currently based on marked point process for geometric feature extraction from high resolution remote sensing images, a marked point process is limited to a single type of objects with simple geometric shape. Moreover, the complexity of interactions between the objects defined in the model makes it impossible to generalize each particular model to another application..In order to extend the level of generality, this project proposes a new stochastic marked point process with irregular polygons as marks for describing geometrical features in a high spatial resolution remote sensing image. Based on the proposed marked point process, a stochastic model for high spatial resolution remote sensing image is built, which allows the representation of images in terms of complex geometry shape rather than a regular geometry. After a probability distribution measuring the quality of geometric feature is specified, the maximum density estimator is searched for by RJMCMC scheme. It can be anticipated that the approach to geometric feature extraction developed with the proposed marked point process is more general and works efficiently on various remote sensing-based applications such as classification of land use and cover, urban object extraction, disaster management, environment monitoring, whereas the conventional approaches needed to exploit specialized models for each problem.

高空间分辨率遥感影像提供地球表面丰富的几何信息，从而为精确提取其中的几何特征提供了充分的依据和可能。同时，高空间分辨率也使得遥感影像中同一地物目标内像素光谱测度的相似性减弱，不同地物目标间像素光谱测度的差异性减弱，以及地物目标内几何噪声增大，由此增加了几何特征提取的难度。为了解决上述矛盾，本项目针对高空间分辨率遥感影像几何特征提取问题开展系统的理论与实践研究。立足于随机几何中标识点过程的理论与方法，重点研究以非规则多边形为标识的标识点过程的构建及特性、地物目标几何形状的非规则多边形拟合、几何特征融入图像建模等问题。在此基础上建立基于地物目标本身而非像素的几何特征提取模型，为开发具有广泛适用性的基于标识点过程的高空间分辨率遥感影像几何特征提取算法奠定坚实基础。研究成果将在基于高空间分辨率遥感影像的大规模土地利用/覆盖分类、城市目标提取、灾害评估、环境监测等方面发挥作用。

当前，遥感学界面临的最具挑战性任务之一是设计行之有效的数据处理方法，以适应日益提高的遥感数据空间分辨率。随着空间分辨率的提高，遥感数据中地物目标呈现更加精确的几何结构和更加细致的细节特征，为地物目标的精准解译奠定了数据基础，但同时亦为其带来更大的难度。首先，需要考虑的问题是如何有效地建模地物目标几何结构，使之适用于地物目标提取要求；其次，空间分辨率的提高在减弱地物目标内观测数据同质性的同时，亦减弱了地物目标间、地物目标与其背景间观测数据异质性，极大地增加了数据建模的难度。为了解决上述问题，项目提出结合点过程和统计建模的地物目标提取策略。在此思想指导下重点研究：（1）利用点过程理论建模地物目标几何结构：分别以非规则多边形为标识的标识点过程和聚类点过程中的聚集点刻画地物目标的空间分布及其几何形态；（2）建立结合几何结构及观测数据的地物目标提取模型：利用地物目标内观测数据的统计一致性、地物目标与背景环境间观测数据的统计差异性等建立地物目标数据模型，并结合几何和数据模型建立地物目标提取模型；（3）设计地物目标提取模型的优化模拟算法：设计实现基于可逆变马尔可夫链蒙特卡洛模型以及基于朗之万方程的跳变-扩散模型的两类优化模拟算法；（4）提出利用地物目标几何形态外延区定义的一、二类误差测度，以及提取和参考几何形态间的相似性测度，以评价算法精度。根据提出的地物目标提取方法学设计实现：基于规则和非规则几何标识的LiDAR点云建筑物提取算法、以及基于非规则几何标识的SAR影像海洋溢油区提取算法，从而验证了提出方法学的可应用性和有效性。提出的针对高分辨率遥感影像的地物目标提取策略，不仅给出一种广普的表征任意面状地物目标几何形态的点过程方法，并将其融入高分辨率遥感影像地物目标提取模型中，而且该思想亦可扩展到建模具有其它几何特征（如线状、网状结构）的地物目标，以及其它高分辨率遥感影像处理任务中（如影像分割）。