遥感图像森林信息的膨胀-剔除提取方法研究

Based on the expanding-eliminating method( EEM), which was proposed by the proposer of this program in former researches, this program will take a systematical research to develop a totally new method of extracting forest information from remotely sensed imagery.This new method will have no limits of number of features and number of target land covers. Research includes: 1, feature selection and feature compression; 2, selecting 3 features from the total for display in an image for a given target lang cover and enhancing of the image; 3. development of EEM algorithm based the high multi-dimengsion chain-style histogram. The basic course is: choosing a study sample to forming seed classes; expanding the seed classes in the high-dimension histogram based the mathematical morphology method, and non-target class grids are permitted to enter the target classes; using the mouse to extract the pixels that they might be mistakenly classed to the target classes in the expanding course on the image, eliminating the histogram grids where the wrong pixels belong to from the target classes and they are not allowed to enter the target class, which was mis-entered in former expanding operation, again. Repeat the expanding-eliminating process until a satisfactory result occurs. The nature of this new EEM includes: the image space and the feature space are in a close linkage as the processing course goes on; suppervised in the whole course and convenience for man-machince interactive operation; is totally a non-parametric method and has a high data-adaptability.

申请者在以往研究中提出了"膨胀-剔除信息提取法"，但这个方法只适用于3维特征图像和单目标地类情况。本研究进行拓展性研究，使之突破特征数量和目标地类数量的限制，从而适用于一般遥感图像的森林信息提取。为了实现这些突破，进行如下内容的研究：分类特征选择与特征提取，去除冗余信息；从众多特征中选择针对某一目标地类的用于图像显示的3个特征，以及在选出的3个特征基础上进行可视化增强；基于高维链表式直方图的膨胀-剔除算法。方法的基本过程是：选取学习样本组成种子类；在直方图内基于种子类进行数学形态学膨胀，允许有非目标区域像素的对应栅格进入目标类；用鼠标在图像上提取错误膨胀区域像素，将这些像素所在的栅格从直方图目标类中剔除，并规定这些栅格在以后的膨胀操作中不被再次膨胀回以前错误膨胀的地类。反复进行膨胀和剔除操作，直至得到满意。该方法的特色是特征空间与图像空间有机联动；全程监督，全程直观；完全非线性。

项目的研究重点在于：分类特征选择与特征提取，去除冗余信息，从众多特征中选择或提取针对某一目标地类的用于图像显示的3个特征。项目开发了以下方法：（1）基于多元线性回归的特征极化选择方法Ⅰ，在最小二乘法或无偏的岭回归方法获得的参数基础上，通过惩罚函数来极化参数，即使得部分参数趋向于零，从而达到选择参数的目的，惩罚函数是本项目提出来的，是相乘形式的，计算例子表明可以取得较好的效果，缺点是当特征量非常大的时候，惩罚函数有溢出的风险；（2）基于多元线性回归的特征极化选择方法Ⅱ，为避免惩罚函数溢出的风险，将惩罚函数改写为对数形式，使得计算更为稳定可靠；（3）将无骗的岭回归方法应用到NNG方法，使其不受变量共线性和特征数大于样本容量问题的影响；（4）改进了以前项目负责人提出的线性回归变换法，将特征按大小进行排序，建立特征与顺序号之间的直线回归方程，除了截距和斜率作为特征外，将决定系数也作为特征，实现将众多特征变换到3个特征的目的。有了特征选择和提取方法作支撑，就可以更好地使用膨胀-剔除信息提取方法。该方法的特色是特征空间与图像空间有机联动；全程监督，全程直观；完全非参数。