ABSTRACT

Relative radiometric normalization (RRN) is a critical preprocessing step that is widely applied to remote sensing data. Essential to RRN are the pseudo-invariant features (PIFs) – ground objects with invariant reflectance across the acquisition dates of multiple images. To meet the requirements of RRN for change detection in multi-temporal images, this paper proposed a novel automatic method for PIF extraction: first, we calculated spectral, spatial, and visual features for the images. Next, a set of change vectors were generated to quantify the features’ differences. Then, the initial PIF was obtained by synthesizing similar pixels from each change vector. Finally, an iterative linear regression was performed to generate the final PIFs. Furthermore, we employed the random sample consensus (RANSAC) algorithm to establish the RRN model for normalizing the target image. The methods and workflow of PIF extraction and the RRN model establishment were validated with two TripleSat-2 multi-spectral images with a spatial resolution of 3.2 metres. The applicability of the proposed PIF extraction method was further validated by Landsat-8 OLI, GF-1 WFV, Sentinel-2 MSI, and WorldView-2 images. The results showed that: (1) The proposed PIF extraction method can automatically obtain the PIFs with high precision. Based on the extracted PIFs, the RANSAC algorithm is capable of building the RRN model with a significant linear relationship. (2) The PIF extraction method is applicable to multi-spectral images with different spatial resolutions. (3) Compared with the multivariate alteration detection (MAD) and the iteratively reweighted MAD (IR-MAD), the proposed PIF extraction method achieved a better performance in selecting the pixels with smaller radiometric differences, which is helpful to reach a higher accuracy of the RRN model. (4) Differences in image brightness, spectral domain, and spatial domain are considered in the PIF extraction method with fewer parameters and high portability. (5) Spectral features are dominant to obtain ideal PIFs, and feature combinations can get the advantages to enhance the effectiveness of the PIFs.

摘要

相对辐射归一化（RRN）是一个关键的预处理步骤，广泛应用于遥感数据。RRN 必不可少的是伪不变特征 (PIF)——在多幅图像的采集日期内具有不变反射率的地面物体。为了满足RRN对多时相图像变化检测的要求，本文提出了一种新的PIF自动提取方法：首先，我们计算了图像的光谱、空间和视觉特征。接下来，生成一组变化向量来量化特征的差异。然后，通过从每个变化向量合成相似的像素来获得初始 PIF。最后，执行迭代线性回归以生成最终的 PIF。此外，我们采用随机样本一致性（RANSAC）算法来建立用于归一化目标图像的 RRN 模型。PIF 提取和 RRN 模型建立的方法和工作流程通过两幅空间分辨率为 3.2 米的 TripleSat-2 多光谱图像进行验证。Landsat-8 OLI、GF-1 WFV、Sentinel-2 MSI 和 WorldView-2 图像进一步验证了所提出的 PIF 提取方法的适用性。结果表明：(1)提出的PIF提取方法可以自动获得高精度的PIF。基于提取的 PIF，RANSAC 算法能够构建具有显着线性关系的 RRN 模型。(2) PIF提取方法适用于不同空间分辨率的多光谱图像。(3) 与多变量变化检测(MAD)和迭代重加权MAD(IR-MAD)相比，所提出的PIF提取方法在选择辐射差异较小的像素方面取得了更好的性能，有助于达到更高的精度。 RRN 模型。(4) PIF提取方法考虑了图像亮度、谱域和空间域的差异，参数少，可移植性高。(5) 光谱特征占优势以获得理想的PIF，特征组合可以获得优势以提高PIF的有效性。(4) PIF提取方法考虑了图像亮度、谱域和空间域的差异，参数少，可移植性高。(5) 光谱特征占优势以获得理想的PIF，特征组合可以获得优势以提高PIF的有效性。(4) PIF提取方法考虑了图像亮度、谱域和空间域的差异，参数少，可移植性高。(5) 光谱特征占优势以获得理想的PIF，特征组合可以获得优势以提高PIF的有效性。