Pseudo Invariant Features Selection for Optical Satellite Images Using Multitemporal and Multivariate Alteration Detection

• Main Authors: WangChih-Chia, 王志嘉
• Other Authors: Chao-Hung Lin
• Format: Others
• Language: en_US
• Published: 2017
• Online Access: http://ndltd.ncl.edu.tw/handle/hav628

碩士 === 國立成功大學 === 測量及空間資訊學系 === 105 === Radiometric normalization is a fundamental preprocessing for multitemporal optical satellite images. The methods of radiometric normalization can be classified into absolute and relative normalization based on the data required in the algorithm. Absolute normalization converts image digital numbers to Earth surface reflectance with the aids of sensor calibration data, atmospheric correction model, and sun angle, which are not always available. In contrast, relative normalization converts digital numbers of subject images to that of a selected reference image or to a common reference domain without the requirement of additional data. However, the accuracy of relative normalization depends on the quality of selected Pseudo Invariant Features (PIFs). PIFs represent the ground objects whose reflectance are constant during a period of time. In previous study, a method, called Multivariate Alteration Detection (MAD), was applied to statistically select no-changed pixels in bi-temporal satellite images. However, MAD is sensitive to significant land-cover changes such as cloud covers. Several clouds may be misclassified as PIFs in this method. For this reason, Iteratively Reweighted MAD (IR-MAD) was introduced to establish a better no-changed background using iterative scheme. Nonetheless, both MAD and IR-MAD compute linear combinations which are suitable for bi-temporal images only, and are not applicable for multitemporal images with more than two images. In this study, a novel method called Weighted Generalized Canonical Correlation Analysis (WGCCA) is proposed for the selection of high-quality PIFs for multitemporal and multispectral images. The proposed method computes correlation coefficients for not only multivariable data but also multitemporal data. Specifically, the method integrates the strengths of Generalized Canonical Correlation Analysis (GCCA) and IR-MAD, and PIFs are extracted simultaneously from a sequence of satellite images, which leads to a consistent PIFs extraction. Furthermore, when the high-quality PIFs are determined by the proposed method, the digital numbers of PIFs from multitemporal images are transformed into a predefined radiometric reference level. With this approach, the radiometric resolution of multitemporal images can be preserved. In experiments, SPOT-5 imagery was tested. Compared with Canonical Correlation Analysis (CCA) which is used in MAD and IR-MAD, the proposed method can discriminate no-changed pixels from changed more accurately.