DEM GENERATION FROM HIGH RESOLUTION SATELLITE IMAGES THROUGH A NEW 3D LEAST SQUARES MATCHING ALGORITHM

International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume XXXVIII-4/W19, 2011 ISPRS Hannover 2011 Workshop, 14-17 June 2011, Hannover, Germany DEM GENERATION FROM HIGH RESOLUTION SATELLITE IMAGES THROUGH A NEW 3D LEAST SQUARES MATCHING ALGORITHM Taejung Kim, Sooahm Rhee Department of Geoinformatic Engineering, Inha University, 253 Yonghyun-Dong, Namgu, Incheon Korea , Commission I, WG I/4 KEY WORDS: Matching, DEM, Least Squares Estimation, High Resolution Satellite images ABSTRACT: Automated generation of digital elevation models (DEMs) from high resolution satellite images (HRSIs) has been an active research topic for many years. However, stereo matching of HRSIs, in particular based on image-space search, is still difficult due to occlusions and building facades within them. Object-space matching schemes, proposed to overcome these problem, often are very time consuming and critical to the dimensions of voxels. In this paper, we tried a new least square matching (LSM) algorithm that works in a 3D object space. The algorithm starts with an initial height value on one location of the object space. From this 3D point, the left and right image points are projected. The true height is calculated by iterative least squares estimation based on the grey level differences between the left and right patches centred on the projected left and right points. We tested the 3D LSM to the Worldview images over ‘Terrassa Sud’ provided by the ISPRS WG I/4. We also compared the performance of the 3D LSM with the correlation matching based on 2D image space and the correlation matching based on 3D object space. The accuracy of the DEM from each method was analysed against the ground truth. Test results showed that 3D LSM offers more accurate DEMs over the conventional matching algorithms. Results also showed that 3D LSM is sensitive to the accuracy of initial height value to start the estimation. We combined the 3D COM and 3D LSM for accurate and robust DEM generation from HRSIs. The major contribution of this paper is that we proposed and validated that LSM can be applied to object space and that the combination of 3D correlation and 3D LSM can be a good solution for automated DEM generation from HRSIs. We believe that the reason for the failure of producing DEMs of good quality lies in matching schemes. Most of stereo matching algorithms proposed for HRSIs are based on image-space search. For an image point within a reference image, its corresponding image point within a target image is searched for (Lee et al., 2003). This scheme may work well with medium resolution images, where the singularities are rarely observed. This scheme, however, is vulnerable for HRSIs, where image points often lie on the singularities. To overcome this problem, we believe that matching has to be performed in an object space with a more intelligent correspondence search scheme. 1. INTRODUCTION Understanding the exact shape of the earth surface has been one of the primary goals of the modern science. It will enable precise biophysical study of the earth environments and accurate analysis of human effects on the earth environments. Very powerful sources to capture the earth surface shape are high resolution satellite images (HRSIs). They are free from accessibility and can be obtained globally consistently. They offer cost-effective and visible data to model the earth surface. Hence, automated generation of DEMs from HRSIs has been an active research topic for many years. On the other hands, object-space matching schemes define voxels within a 3D object space, project each voxel onto the left and right image spaces, and check similarities by grey-level correlation. Often they are very time consuming and critical to the dimensions of voxels. This method may not produce better performance to the matching algorithms based on the image space. It is known that stereo matching of HRSIs is difficult due to singularities caused by the intrinsic nature of them, such as occlusions and building facades (Kim, 2005). Height profiles generated from HRSIs are often very crude and contain blunders or holes. In many occasions there is little improvement in DEMs from HRSIs compared to those from medium resolution images. For this reason multiple optical images (Okutomi and Kanade, 1996; Gruen and Baltsavias, 1998), instead of two, or other data sources, such as laser scanner data and differential SAR data, are utilized for DEM generations. In this paper, we tried a new least square matching (LSM) algorithm that works in a 3D object space. The algorithm starts with an initial height value on one location of the object space. From this 3D point, the left and right image points are projected. The true height is calculated by iterative least squares estimation based on the grey levels of the left and right patches centred on the projected left and right points. This 3D LSM can overcome the problems of the image space matching since the projected image points will not lie on the singularities. The 3D LSM will overcome the problems of the conventional object space matching algorithms since there is no need to define the voxel dimensions. In this paper, we report our work on automated DEM generation from stereo pairs of HRSIs. We are pursuing our research under the activities of the ISPRS WG I/4 on ‘benchmarking and quality analysis of DEM generated from high and very high resolution optical stereo satellite data’ and we report our interim results. We propose a new matching scheme for automated DEM generation that utilizes least squares correlation matching in an object space. We compare the new matching scheme with existing ones and combine them together for reliable DEM extraction. 153 International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume XXXVIII-4/W19, 2011 ISPRS Hannover 2011 Workshop, 14-17 June 2011, Hannover, Germany We tested the 3D LSM to the Worldview images over ‘Terrassa Sud’ provided by the ISPRS WG I/4. We compared the performance of the 3D LSM with the correlation matching based on 2D image space, (Lee et al., 2003) and a correlation matching based on 3D object space. The accuracy of the DEM from each method was analysed against the ground truth. Test results showed that 3D LSM offers more accurate DEMs over the conventional matching algorithms. Results also showed that 3D LSM is sensitive to the accuracy of initial height value to start the estimation. We combined the 3D correlation matching and 3D LSM for accurate and robust DEM generation from HRSIs. According to equation (1), we can compute the differential equations of image coordinate by equation (4). (4) A combination of equation (3)-(4) yields (5) 2. 3D LEASET SQUARES MATCHING METHOD This section describes a new least square matching method based on the 3D object space, 3D LSM. The LSM based on the 2D image space (2D LSM) is a well-known method to prod (...truncated)