摘要:
为了提升农田高分遥感影像(high spatial resolution remote sensing image,HRI)的信息提取效果,提出了一种新的农田HRI分割算法.传统的Mean-Shift(MS)HRI分割算法仅利用全局或单一的尺度参数;而常规可变尺度MS算法在尺度参数估算中也只考虑光谱信息.这些都导致其分割结果难以完整地展现不同尺度的农田区域.针对该问题,在MS算法的基础上进行了改进: 第一,提出了一种局部可变尺度参数的估计方法;第二,提出了利用局部可变尺度进行MS滤波的模型函数.该改进算法主要包含3步: ①为了全面考虑不同波段的响应变化,在MS滤波核函数中采用了对角化的尺度参数矩阵,并将其与采样点密度估计模型相结合,导出了一种可变尺度MS滤波的迭代函数;②为了提高算法的自动化程度,利用局部光谱变化与边界强度信息,提出了一种新的局部尺度参数估算方法;③将MS滤波结果输入到基于分形网络演化方法(fractal net evolution approach,FNEA)的空间聚类算法中,得到最终的分割结果.利用RapidEye与OrbView3的2景HRI进行了算法验证.实验结果表明,所提出的改进算法能够优化农田HRI分割的精度.
Abstract:
In order to improve the effect of information extraction from high spatial resolution remote sensing images (HRI) of cropland, the authors put forward a new HRI segmentation algorithm.Due to the fact that the traditional Mean-Shift (MS) segmentation method only uses a global and single scale, and that some variable bandwidth MS only considers spectral information in their scale estimation process, and croplands with various sizes could be hardly extracted in one segmentation result, the authors improved a MS based approach to tackle this problem.The main consideration lies in two aspects: ① A local variable scale parameter estimation method is proposed;② The model function for local variable scale is established for MS filtering.The proposed approach mainly consists of 3 parts: ① With the objective of comprehensively considering the response variation of different bands, the diagonal scale parameter matrix is adopted in the kernel function of MS filtering, and it is combined with sample point estimation model to derive the iterative function for variable scale MS filtering;② For the purpose of increasing automation of the proposed method, local spectral variation and edge strength information are utilized to design a new local scale parameter estimation method;③ For obtaining the final segmentation, the filtering result is used as input for the fractal net evolution approach (FNEA) which is a spatial clustering method.Two scenes of HRI acquired by RapidEye and OrbView3 were employed for experiment, and the results show that the proposed method can optimize the accuracy of cropland HRI segmentation.
下载: