CO-REGISTRATION BETWEEN MULTISOURCE REMOTE-SENSING IMAGES

International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume XXXIX-B3, 2012 XXII ISPRS Congress, 25 August – 01 September 2012, Melbourne, Australia CO-REGISTRATION BETWEEN MULTISOURCE REMOTE-SENSING IMAGES Joz Wu a,*, Chi Chang b, Hsien-Yu Tsai b, Ming-Che Liu b a b CSRSR, National Central University, Jhongli, Taiwan – Dept. of Civil Engineering, National Central University, Jhongli, Taiwan – Commission III, WG III/5 KEY WORDS: Registration, Least-Squares Matching, SIFT, TPS, RANSAC ABSTRACT: Image registration is essential for geospatial information systems analysis, which usually involves integrating multitemporal and multispectral datasets from remote optical and radar sensors. An algorithm that deals with feature extraction, keypoint matching, outlier detection and image warping is experimented in this study. The methods currently available in the literature rely on techniques, such as the scale-invariant feature transform, between-edge cost minimization, normalized cross correlation, leastsquares image matching, random sample consensus, iterated data snooping and thin-plate splines. Their basics are highlighted and encoded into a computer program. The test images are excerpts from digital files created by the multispectral SPOT-5 and Formosat2 sensors, and by the panchromatic IKONOS and QuickBird sensors. Suburban areas, housing rooftops, the countryside and hilly plantations are studied. The co-registered images are displayed with block subimages in a criss-cross pattern. Besides the imagery, the registration accuracy is expressed by the root mean square error. Toward the end, this paper also includes a few opinions on issues that are believed to hinder a correct correspondence between diverse images. accepted in disciplines like computer vision, photogrammetry and remote sensing. In the framework of image pyramiding, convolution with blurring Gaussian kernels is carried out first. The images are consecutively differenced to yield a stack of scale-space images containing potential high-pass feature points. 1. INTRODUCTION In general, there are image feature- and area-based matching methods. Linear features can be extracted from a portion of an image where the gray-level gradient varies (Tupin and Roux, 2003). On the other hand, the coefficient of correlation between two image windows may serve as an index for gauging the degree of similarity (Wolf and Dewitt, 2000). A hybrid method allowing for both feature- and area-based matching techniques is considered more versatile than either method operating alone. No doubt, the design of a hybrid strategy could lead to a more complex algorithm with heavy computation. Often, this is a blessed trade-off because of the increased reliability of point determination. Symbol D is used to represent the resulting images as D T 2D 1 ) x  xT x , based on truncated x 2 x 2 Taylor’s expansion with the capital superscript standing for D ( x)  D  ( transposition, and with xT  ( x, y ,  ) , a vector having the line x (pixel), sample y (pixel) and scale  coordinates. For an extreme keypoint, the differentiated equation of D with respect This paper is motivated to devise an algorithm that stresses not only the matching accuracy and robustness between images, but also the scale- and rotation-invariance between them. Many generic methods for feature extraction and image registration exist (Dare and Dowman, 2001; Mikolajczyk and Schmid, 2005). In particular, Lowe (2004) published a scale-invariant feature transforming methodology, which allows us to generate a large number of descriptor-based keypoints. Gross errors in the coordinates of the image keypoints have to be detected and removed, on a probabilistic basis (Schwarz and Kok, 1993; Vennebusch et al., 2009). Indeed, the filtered feature points possess good coordinate approximates. They may serve as initial values for the subsequent high-precision least-squares image matching. to x is 2D x 2 x D 0 . x Consequently, one obtains  2 D 1 D . After back-substitution, the image is ) x x 2 identified as xˆ  ( D(xˆ )  D  1 D T ( ) xˆ 2 x (1) The magnitude and orientation of gray-level gradients in the image closest to a keypoint results in an orientation histogram that accounts for a relative rotation between image windows, within a plus or minus 5-degree tolerance. Based on the aligned image at a keypoint, an SIFT user sets up a vector of 128 descriptive elements. Search for the corresponding point to form a pair relies on a minimization of the Euclidean distance between two descriptive vectors. 2.2 Area-based Matching 2. METHODOLOGY 2.1 Feature Points by SIFT The SIFT (Scale-Invariant Feature Transform) algorithm by Lowe (2004) has been famously known for its insensitivity to imaging scale and orientation changes, and to scene illumination differences, thereby allowing it to be widely Generally speaking, the method of LSM (Least-Squares Matching) outperforms that of normalized cross correlation because the former incorporates affine parameters into the * Corresponding author. This is useful to know for communication with the appropriate person in cases with more than one author. 439 International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume XXXIX-B3, 2012 XXII ISPRS Congress, 25 August – 01 September 2012, Melbourne, Australia blunder is removed one at a time so that the algorithm is termed IDS (Iterated Data Snooping). matching model. For the target g ( x, y ) and search q (l , s ) images, the line and sample coordinates (pixel) are expressed as l  a0  a1x  a2 y and s  b0  b1x  b2 y , with the a0 , a1 , As calculation of the covariance matrix for a large number of data residuals can grow burdensome, the RANSAC algorithm is usually conducted first. For the remaining data points, IDS could be invoked to ensure that indeed they are regular samples of an experiment at hand. Because of the difference in theory, RANSAC and IDS are expected to be complementary. a 2 , b0 , b1 and b2 symbols expressing affinity. Differencing pixel values may lead to a gray-level function as vi  h0  h1qi (a0  a1 x  a 2 y, b0  b1 x  b2 y )  g i ( x, y )  0. Inde x i varies for n pixels in a window. Symbol vi denotes a zeromean residual error having the Gaussian distribution, or 2.4 Thin-plate Spline Interpolation N (0, i2 ) with the  i symbol meaning the standard deviation; TPS (Thin-Plate Splines) stands for a flexible function in that it emulates the minimized bending energy of a metal plate on multiple tie-point constraints. A trend surface stems from a global, affine transformation between two overlapping images. h0 and h1 linearly modify pixel values. Linear expansion at approximate unknowns results in a system of error equations, defined as v  Ax  l with  02Q . By referring to Mikhail (1976), one obtains the least-squares solution of unknown p (...truncated)