Development of Efficient Registration Algorithms for Remote Sensing Optical and SAR Images

Abstract

Over the last two decades, the volume and quality of the remote sensing images have increased tremendously. The extensive availability of the images has led to a significant boost in application areas. The remote sensing images are generally acquired by the earth observing satellites and the airborne-based imaging systems using single or multiple sensors. The passive optical sensors capture the remote sensing optical images whereas the active
synthetic aperture radar (SAR) sensors capture the SAR images. These images are widely used in many applications such as change detection, image mosaicing, and image fusion.
In these applications, image registration is used as a fundamental step which performs a precise image-to-image alignment. However, the remote sensing image registration is
considered to be a challenging task due to the presence of significant geometric as well as radiometric differences between the images and the noise affect. To address these issues, this dissertation proposes different efficient algorithms for optical-to-optical, SAR-to-SAR,and optical-to-SAR image registration.

Due to its robustness in orientation, scaling, and view-angle differences, scale-invariant feature transform (SIFT) has been widely used for optical-to-optical image registration. However, it sufferers from lack of correct matches and uneven distribution of the matching
pairs which affect the performance of registration. In order to improve the distribution quality of the matching pairs, a modified uniform robust SIFT (M UR-SIFT) algorithm is
proposed. In addition, a final matched feature selection(FMFS) algorithm is presented to increase the number of correct matches.

Most of the existing SAR-to-SAR image registration methods suffer from lack of controllability of the number of extracted features and uneven distribution of the matching
pairs. Moreover, the presence of multiplicative speckle noise in SAR images further affects the performance of feature extraction and matching methods. To address these problems, three different SAR-to-SAR image registration algorithms are proposed. The first algorithm is an improved version of the synthetic aperture radar-scale invariant feature transform (I-SAR-SIFT). It can extract robust, distinctive, and uniformly distributed features from
the SAR images. The second one is a block-based multi-feature extraction (BBMFE) algorithm. In this method, two types of features are extracted from the SAR images to
increase the number of extracted features and to improve the distribution quality of the matches. The influence of speckle noise reduces the distinctiveness of the SIFT-based
descriptor. Therefore, the third SAR-to-SAR image registration algorithm is developed to improve the distinctiveness of the SIFT-based descriptor. A Gabor odd filter-based ratio operator (GOFRO) is utilized to construct the feature descriptor in our proposed method
because it has comparatively better robustness in the presence of speckle noise than the other ratio operators.

Generally, the optical and SAR images have significant radiometric as well as geometric differences. To address these issues, an optical-to-SAR image registration method is
proposed. The proposed method uses a structural descriptor to register the optical and SAR images having significant geometric differences. It can increase the number of correct
matches as well as correct matching rate in optical-to-SAR image registration. The performance of the proposed methods is evaluated by using different remote sensing image pairs.