高分遥感图像相对辐射校正中的伪不变地物自动提取和优化选择

相对辐射校正是遥感变化检测中重要的预处理过程,伪不变地物（Pseudo-Invariant Features,PIF）是多时相影像中相对不变的地物,是相对辐射校正中的重要依据。针对高分遥感图像变化检测中相对辐射校正的要求,本文提出了一个自动提取和优化选择PIF的流程和方法：首先计算两期图像的亮度、光谱特征和空间特征的变化向量,然后对各变化向量的像元值从低到高进行排序,经多数投票后提取PIF,最后使用“迭代线性回归—去除异常值”方法选择获得最终PIF。以2016年11月27日和2017年7月18日的2期“北京二号”高空间分辨率多光谱影像为例,选择地物占比不同的两个实验区对流程和方法进行了验证,并与多元变化检测和迭代加权多元变化检测的PIF提取方法进行了比较。使用两期WorldView-2影像和Landsat-8 OLI影像对方法的适用性进行了验证。结果表明：① 2个实验区提取的PIF精度分别为98.74%和98.71%,PIF像元合理分布于未变化区域、包括了影像中主要的地物类型;② 使用本文方法提取的PIF建立的相对辐射校正模型具有显著的线性拟合效果（p<0.000 1）;③ 本文方法考虑了图像亮度、光谱信息以及空间信息的差异,使用参数少,可操作性高;④ 与多元变化检测和迭代加权多元变化检测方法相比,本文方法提取的PIF更为合理,建立的辐射校正方程拟合效果更佳;⑤ 本文方法适用于具有相同波段设置的中、高空间分辨率光学遥感影像。

Automatic Extraction and Optimal Selection of the Pseudo Invariant Features for the Relative Radiometric Normalization in High-Resolution Remote Sensing Imagery

Relative radiometric normalization is an important process in the remote sensing image change detection. Identifying the Pseudo-Invariant Features (PIF) with invariant or near-invariant radiometric reflectance over a certain period in multi-temporal images is a key to radiometric normalization. This paper proposed a novel method for automatic extraction and optimal selection of PIF. First, the change vectors including brightness, and spectral and spatial domains of bitemporal images were generated. Then, the pixels of each change vector were sorted from the lowest to the highest value, and the majority vote algorithm was used to extract the initial PIF. Finally, the PIF was selected by the iterative linear regression and outlier analysis. Taking two multi-spectral high-resolution Tripesat-2 images acquired on November 27, 2016 and July 18, 2017 as example, two typical regions with different land cover types were selected to test the proposed method. The proposed method was compared with Multivariate Alteration Detection (MAD) and Iteratively Reweighted Multivariate Alteration Detection (IR-MAD) methods. The applicability of the proposed method was also validated by two WorldView-2 and Landsat OLI images. The results show that: (1) the accuracy of the PIF that extracted within two regions were 98.74% and 98.71%, respectively. The extracted PIF was distributed in the unchanged areas and covered the main land cover types in the images; (2) the linear regression models of the relative radiometric normalization using the PIF extracted by the proposed method were significant (p<0.000 1); (3) the differences in image brightness, spectral domain, and spatial domain were taken into account in this method with less parameters and high operability; (4) compared with MAD and IR-MAD, the proposed method showed a better performance in extraction precision and an significant linear regression model of the relative radiometric normalization; and (5) the proposed method was suitable for other medium- or high-resolution remote sensing images with same bands.