The effects of incidence angle, topography and land cover on automated SIFT based orthorectification of VHR satellite images

Developments in satellites with optical satellites have been evolving rapidly. Recent developments enabled end users from different disciplines such as urban mapping, transportation network determination, presicion agriculture and homeland security to conduct their researches using satellite images. Remotely sensed images contain geometric distortions related to physical parameters such as earth curvature and rotation, surface topography, orbit of the satellite and incidence angle. These errors related to the physical parameters, limit satellite images to be used in analysis directly. Thus, it is an important step to geometrically correct satellite images before conducting analysis. The relationship between satellite images and surface can be defined using mathematical models as well as it can be defined using physical parameters, which are interior and exterior orientation parameters. Rational Functional Model (RFM) is one of the mathematical models used for geometric correction and the required Rational Polynonial Coefficients (RPC) to set these functions are usually delivered to end user with the satellite image. Geometric correction process operated using RFM, which does not require Ground Control Points (GCP), can eleminate some of the errors related to physical parameters but some random errors and bias remains in the geometrically corrected image. Several researches show that RFM refinement method using GCPs and Digital Elevation Model (DEM) increase the spatial acuracy and can be considered as better option for geometric correction than using original RFM model. RFM refinement methods are grouped in two categories namely, direct and indirect methods. Direct RFM refinement methods use the existing parameters which are already used for RPC generation along with the new GCPs measured from the reference data. On the other hand, indirect refinement approch use only the new GCPs to achieve better results. While direct refinement method calculate RPC coefficients from scratch, indirect refinement method only refines the end result obtained from RFM model which uses original RPC. In conventional methods, GCPs which are required for RFM refinement are manually measured by remote sensing operator using reference data having high spatial accuracy. When large areas are considered for geometric correction, as remote sensing operators are involved in, much more workforce and time will be required and the accuracy of the related work will be related directly with the operator's skill and experience. To overcome the problems related to workforce and time requirement, image matching algorithms have been developed to help geometric correction proccesses operate autonomously. With the help of image matching algorithms, GCPs are generated with matching source satellite images with reference data having high spatial accuracy. Since these methods are creating eneromous number of GCPs, they may also contain mismatches. Results of the matching process are filtered as they may have mismatched points using RANSAC method and remaining matching points which pass the filtering operations are saved as GCPs. Adoption of the image matching algorithms have been increasing day by day to complete geometric correction operations faster and minimize the errors related with the operators thanks to rapid developments in the image matching algorithms and computer hardware. Scale Invariant Feature Transform (SIFT) algorithm is one of the matching algorithms utilized by more resarchers in remote sensing domain thanks to rapid developments in computer hardware. SIFT is providing scale and rotation invariance which is usually perfect condition to use this algorithm with remote sensing data as scale and rotation parameters may vary in different satellite images which are acquiered by different sensor. Since, SIFT is a computer vision method, spatial accuracies of the GCPs generated using SIFT algorithm may vary in different conditions. Reasons for differences in spatial accuracies can be explained in terms of differences in incidence angle, surface topography or land cover parameters. In this research, effects of the incidence angle, surface topography and land cover parameters to the SIFT based autonomous geometric correction process are analyzed. Analyses have been conducted on three different study sites, which are inside of the borders of Istanbul, Bursa and Izmır cities respectively, having different characteristics using Pleiades satelite images. Two images having low incidence angle and one image having high incidence angle were selected for this study. One of the images whose incidence angle close to 0 in the low incidence angle dataset selected as reference satellite image. Reference data has been created using GCPs manually measured from orthophotos having 1:5000 scale. Slope maps have been created and analyzed using Digital Elevation Model for each study site. Land Cover maps are directly retrieved from CORINE dataset. RFM refinement model have been selected as a geometric correction model for this study. Direct refinement approach has been utilized to eleminate errors related with the platform itself and biases using 2nd degree polynomial. GCPs required for RFM refinement model have been generated using SIFT based image matching process. Geometric correction process has been conducted for panchromatic channel and for mutlispectral channels of pansharpened image seperately. As researchers ussualy prefer the pansharpened images for analysis purposes, RGB/3 and RGBN/4 data, which was retrieved using aritmethic operation, have been also analysed and evaluated with geometric correction process. Parallel programming approach have been used in geometric correction algorithm development to minimze the required time to complete the operation. The corner coordinates of the source image are retrieved from the metadata and used for core components of the parallel architecture. Results of the study show that the spatial accuracy and the number of the GCPs generated using SIFT image mathching algorithm increases while the incidence angle decrease. Moreover, slope analysis show that there is no clear correlation between SIFT based ortorectificataion results and slope classes. Land cover based analysis show that GCPs generated using SIFT are maninly concentrated in Forest and Semi Natural Areas class. Parallel programming approach helped system to operate 16 time faster than the progress conducted using only one thread.