GEOREFERENCING FROM ORTHORECTIFIED AND NON-ORTHORECTIFIED HIGH-RESOLUTION SATELLITE IMAGERY

Transcription

1 GEOREFERENCING FROM ORTHORECTIFIED AND NON-ORTHORECTIFIED HIGH-RESOLUTION SATELLITE IMAGERY J. Willneff and J. Poon CRC for Spatial Information, University of Melbourne, VIC 3010, Australia, KEY WORDS: Georeferencing, Orthoimage, High resolution, Imagery, IKONOS, QuickBird, Analysis, Accuracy ABSTRACT: The extraction of metric information from image data is highly relevant for the update of geospatial data bases. Photogrammetric methods offer tools to extract such information in a reliable and accurate manner from mostly multiple image setups using appropriate sensor models. These methods require a certain amount of expertise and often expensive software systems to process the image data. An alternative to extract metric information from a multi-image configuration is feature extraction from single, already georeferenced images. The accuracy potential from georeferencing may be sufficient for various applications. This paper describes the method of georeferencing from orthorectified and non-orthorectified high-resolution satellite imagery. During the orthorectification process, georeferencing information enabling the transformation from pixel coordinates to two-dimensional object space coordinates is generated. Mathematically this georeferencing information often supplied in the commonly used Tiff World File (TFW) format describes an affine transformation modelling a two-dimensional shift, two scale factors and two rotations. This additional information is often part of an orthorectified image product. Alternatively, the georeferencing information can also be generated with the use of ground control points (GCPs) and can be applied to both orthorectified and non-orthorectified images. For the generation of a set of the six affine transformation parameters, a minimum of three GCPs is necessary. If only two GCPs are available, a two-dimensional conformal transformation using four parameters can be applied for the image to object space transformation. The paper analyses the two methods of coarse georeferencing applied to orthorectified and non-orthorectified highresolution satellite imagery on the basis of three different data sets and different projection systems. The data processing was performed with the software package Barista. The paper quantifies the accuracy yielded with georeferencing, discusses its limitations and concludes with comments about its suitability for potential applications. 1. INTRODUCTION Images acquired from aerial and space borne platforms are widely used for the extraction of metric 3D information for manifold applications based on geospatial databases. While the processing of such image data via photogrammetric means requires a high level of expertise and often expensive software systems, resulting image products may enable the non-expert user to extract georeferenced information from a single image. Georeferenced images allow 2D object space coordinates determination from pixel positions measured on the image. Generally photogrammetric methods require at least two images with an appropriate sensor model to extract 3D coordinates from 2D image coordinates. This is different in the case of an orthoimage, which is geometrically transformed to uniform scale and where image displacements caused by sensor orientation and terrain relief are removed. The transformation of an orthoimage from pixel to object space coordinates is a 2D affine transformation. The parameters of this affine transformation are calculated during the orthorectification process and often forms part of the orthoimage itself. The set of transformation parameters is either integrated in the header of the orthoimage or come as a separate Tiff World File, or TFW (ESRI, 2005). Commercial GIS software packages allow import of georeferenced image data. There are numerous image formats describing georeferencing information, the most common is the GeoTiff format (Ritter and Ruth, 1997). The transformation parameters model a two-dimensional shift, two scale factors and two rotations. As the transformation describes only the relation between 2D image coordinates and 2D object coordinates, it is therefore not possible to obtain a height value from a georeferenced image. The orthorectification process requires an appropriate sensor model, as well as a digital terrain model (DEM). An overview of digital orthoimagery and the use of orthoimages in GIS are given in (Baltsavias, 1996). If no sensor model information or DEM is available, georeferencing can also be generated with the use of GCPs. The measured image coordinates and the 2D positions of the GCPs allows parameters of 2D transformations, such as affine or conformal, to be determined. At least two GCPs are required to solve the four parameters of a conformal transformation describing a two-dimensional shift, one scale and one rotation. This type of georeferencing is limited in coarse results as the influence of the sensor model and especially the terrain information is completely neglected, although for some applications the accuracy may be sufficient. For instance, a local area with minimal height differences can result in georeferencing of acceptable quality. In the framework of the Cooperative Research Centre of Spatial Information Project 2.1 Automated Mapping & Feature Extraction from Space, Aerial & Terrestrial Imagery, conformal and affine methods of georeferencing were implemented and tested. The data processing was performed with the software package Barista, the project s development platform. Barista is mainly designed for the processing of high-resolution satellite imagery and extraction of 3D information from single and multiple images. This paper describes georeferencing of

2 orthorectified and non-orthorectified image data on the basis of various data sets and the use of different projection systems. It also quantifies the method s accuracy and discusses its limitations. The main goal of the paper is to assess the orthorectification method implemented in the Barista software package, and additionally evaluate a commercially available georeferenced image product. Conclusions are made about the applicability of the coarse georeferencing implemented in Barista. 2. IMAGE DATA SETS The georeferencing was applied to different data sets, both rectified and non-rectified images. This section gives brief descriptions of the data used in the analysis of the georeferencing method. 2.1 Hobart In the Hobart testfield, a total five images were used; two original IKONOS Geo image products supplied by Space Imaging and three images resulting from an orthorectification performed with Barista. In addition to RPC parameters provided with the images, 138 well distributed GCPs and DEM coverage of the area was available. The georeferencing information of the orthoimages was defined in UTM and the coarse georeferencing was performed in UTM. The images cover an area of approximately 13 km x 15 km. This test field is also described in (Fraser and Hanley, 2005). 2.2 Bhutan A stereo pair of QuickBird images, both panchromatic and multispectral was available for the Bhutan testfield. RPC data, a limited number of 12 GCPs and DEM data was available for the area of investigation. A resulting orthoimage with 1 m ground sampling distance (GSD) was generated with georeferencing in the Transverse Mercator (TM) projection system. The georeferencing was also performed in UTM and geographic coordinates. 2.3 Moreton Island A single QuickBird Ortho Ready Standard colour image of Moreton Island with georeferencing information, RPCs and a set of 42 well-distributed GCPs were available for the analyses of the Moreton Island data set. The georeferencing information for this data set is defined in UTM. The image covers an area of approximately 40 km x 12 km and has a GSD of 2.4 m. The Ortho Ready Standard image product is projected to the average elevation of the terrain covered by the scene (DigitalGlobe, 2006). 3. GEOREFERENCED IMAGE DATA Images and raster data are essential data types in GIS applications. One of the most popular image formats is Tag Image File Format (TIFF), which was originally rather limited for cartographic application (Adobe, 1992). In the early 1990 s an initiative under the leadership of Ed Grissom at Intergraph tried to define an extended TIFF specification, more suited to geospatial requirements. In 1995 a GeoTiFF specification was outlined at a conference hosted by SPOT Image, with representatives from USGS, Intergraph, ESRI, ERDAS, SoftDesk, MapInfo, NASA/JPL and others (Ritter and Ruth, 1997). This specification of GeoTIFF extends the set of TIFF tags provided to describe the georeferencing information associated with the imagery, which may originate from aerial or satellite imaging systems, DEMs, scanned maps, or as a result of GIS analyses. Different coordinate system and projections are supported and the georeferencing information is expressed as a set of parameters. The mathematical model allowing the transformation from pixel coordinates to object space coordinates is a two-dimensional affine transformation. Six parameters describe the transformation, expressed in equation 1. X = scale x + rotation y + X geo x pixel x pixel offset Y = rotation x + scale y + Y geo y pixel y pixel offset where x pixel, y pixel = image coordinates X geo, Y geo = coordinates of georeferenced point If the georeferencing information is stored in a World file (e.g. *.tfw, *.jgw, *.pgw, *.bpw, etc.) only six transformation parameters are provided, whereas GeoTIFF imagery may also contain information about the projection system or datum details. In this paper, the georeferencing information is created either during the process of an orthogeneration, or by an affine or conformal 2D transformation using GCPs. In the case of the orthogenerated georeferencing information, the two rotation parameters are 0 as the original image was transformed to the geometry of an unrotated map sheet. 3.1 Georeferencing of ortho-rectified images Orthorectification is the geometric transformation of an image in which image displacements due to sensor orientation and terrain are corrected to the projection of a map coordinate system. The input data for this processing is the original image, an appropriate sensor model and a DEM. In the case of this paper, the input images are high-resolution satellite images from IKONOS and QuickBird provided with RPC (Rational Polynomial Coefficients) sensor models. Original RPCs provided with high-resolution satellite image products are known to contain significant biases, which can be corrected if GCP data of the image area is available (Grodecki and Dial, 2001; Fraser et al., 2006). For the purpose of increased accuracy, the original RPC parameter sets were biascorrected before being used for the orthorectification. This was performed for the Hobart data set. For the Moreton Island data set, a DigitalGlobe Ortho Ready Standard colour image product inclusive of georeferencing information as well as 3D coordinates of GCPs, was made available from SKM for the analyses. 3.2 Georeferencing of non-rectified images If the appropriate sensor model for an image is unknown and accuracy requirements are secondary, coarse georeferencing using either 2D affine or even only a 2D conformal transformation enables the extraction of at least low accuracy information from an image. The requirements are 2D positions of GCPs and the corresponding image coordinates. Using this information, a set of transformation parameters can be calculated within a least squares adjustment enabling the (1)

3 transformation from pixel positions to 2D object coordinates. As neither the influence of the underlying terrain nor sensor model is considered the limitation of this method is rather obvious; although for locally restricted areas or flat terrain, coarse georeferencing may be sufficient for some applications. For this paper, coarse georeferencing was applied to the Hobart data set, which has significant height differences throughout the test area. 4. VALIDATION AND QUALITY ASSESSMENT This section describes the validation procedure for the quality assessment of the different georeferenced data sets. The general idea is the comparison of the 2D position from measured image points in the rectified and non-rectified georeferenced image data with the coordinates of the GCPs. Assuming that the GCPs have a significantly higher accuracy than what can be derived from measurements in the imagery, these coordinates provide a reference against the georeferencing results. As the GCP information is well-distributed over the image area it can be expected to give a profound measure of the overall quality of the georeferencing, giving an indication for the suitability of the different georeferencing approaches. 4.1 Georeferencing applied to data set Hobart The processing of the Hobart data was completely performed with the Barista software. As the original image, RPC parameters, DEM and GCP data was available; the following processing steps were performed: Figure 1. Coordinate differences between GCPs and georeferenced points prior RPC bias-correction; differences displayed with scale factor Optional RPC bias correction via GCPs - Orthorectification with DEM at 5 m post-spacing - Image coordinate measurement of 25 GCPs - Georeferencing via TFW from orthogeneration - Comparison of georeferenced points Figure 1 shows the differences between the georeferenced points and the GCPs. The systematic shift is caused by the bias of the RPC parameters, which was not corrected prior to the orthogeneration. This systematic effect can be removed if the RPC parameters used for the ortho generation are bias-corrected before the ortho generation. Figure 2 shows the differences between georeferenced points and GCPs of an orthorectified image using bias-corrected RPCs. The differences are in the order of one pixel, which corresponds to the accuracy level of the image coordinate measurements. Coarse georeferencing of a non-rectified image was applied to one of the original IKONOS scenes of the Hobart data set. The 138 GCPs were measured in image space and used to determine a coarse georeferencing parameter set. As neither influence of sensor orientation nor terrain is considered, the results can be expected on a rather low level of accuracy, especially with the significant height differences in the terrain imaged on the Hobart scene. Figure 3 shows the results of the coarse georeferencing, huge discrepancies over 100 m in Easting and almost 180 m in Northing between georeferenced points and GCPs occur, especially where the point is well above and below the mean terrain height. Figure 2. Coordinate differences between GCPs and georeferenced points after RPC bias-correction; differences displayed with scale factor 600

4 The points are located in the valley with only little height variation compared to the rest of the surrounding terrain. Due to difficult accessibility in the region, the collection of GCP data was restricted to this limited area. The GCPs were available in three different projection systems. With the control points measured on the orthoimage, it was possible to determine georeferencing parameter sets in UTM, TM and geographic coordinates. Georeferencing was conducted using a 2D affine as well as a 2D conformal transformation. Figure 3. Coordinate differences of coarse georeferencing between GCPs and georeferenced points; differences displayed with scale factor 15 Table 1 shows the differences between georeferenced points and the corresponding GCPs. The accuracy of the orthoimage is significantly improved after the bias correction. The shift in the RPC bias for image 286 was determined to m in Easting and m in Northing. This in turn deteriorates the quality of the resulting orthoimage, where an average shift of m in Easting and m in Northing is determined. Georeferencing method No of RMSE at CKPs [m] GCPs Easting Northing ortho286, TFW without biased RPCs ortho286, TFW after RPC bias correction ortho367, TFW after RPC bias correction Original 286, 2D affine transformation Table 1. Georeferencing applied to Hobart orthoimages 4.2 Georeferencing applied to data set Bhutan The processing of the Bhutan data set included analyses of using georeferencing in different projection systems. An orthoimage with 1m resolution was generated after RPC biascorrection and with the TFW set defined in Transverse Mercator projection. Figure 4 shows the generated orthoimage, which covers an area of roughly 8 km by 6 km, which is only a subset of the original QuickBird satellite scene. The GCPs are located in a limited area of the image, providing a rather sub-optimal distribution. Figure 4. Bhutan orthoimage showing GCPs concentrated in the settled area of Thimpu, marked here within the rectangle Table 2 shows the difference between georeferenced points and GCPs. It shows that georeferencing of an orthorectified image with control points in different coordinate systems is possible on a similar accuracy level as the TFW parameters from the ortho generation. Georeferencing method Coordinate System No of GCPs RMSE at CKPs [m] E or N or TFW ortho generation TM Affine transformation TM Affine transformation UTM Affine transformation Geographic D Conformal TM D Conformal UTM Table 2. Georeferencing in different projections applied to the Bhutan orthoimage

5 The RPC reprojected points show a similar systematic effect as the georeferenced points, although with an even bigger average shift of 19.4 m in Easting and m in Northing. The authors suggest that better georeferencing accuracy could be achieved by applying RPC bias-correction prior to the georeferencing process, but are unable to explain the bigger magnitude of the systematic average shift of the RPC projected points compared to the georeferencing with the TFW parameters. Figure 5. Coordinate differences of coarse georeferencing between GCPs and georeferenced points; differences displayed with scale factor 70 As an orthoimage has generally the geometry of an unrotated map sheet, the 2D conformal transformation should also provide a suitable transformation from image space to 2D object space coordinates. This was applied with the two metric coordinate systems, TM and UTM, producing lower accuracy, but still acceptable results. 4.3 Georeferencing applied to Moreton Island data set The processing of the Moreton Island data set was limited to the image coordinate measurement of the provided GCPs. Not all of the 42 points could be clearly identified in the satellite scene due to the 2.4 m GSD of the colour image. Although some points were not used for the georeferencing analysis due to poor identification, the overall distribution of the points was still covering the whole area of the island. In total 28 points were georeferenced and compared to the UTM coordinates of the ground control data. Table 3 shows the result of the comparison of the georeferenced points against the GCPs. The differences between georeferenced point and ground control show a significant systematic behaviour, with an average shift of 9.3 m in Easting and m in Northing. This may be explained by a bias in the RPC parameters used in the DigitalGlobe georeferencing process. As this image was also provided with RPCs it was possible to project the GCPs to image space using the RPC sensor model. Figure 6 show a visualization of both sets of discrepancies. Georeferencing method No of GCPs Mean discrepancy [m] RMSE at CKPs [m] E N E N TFW DG RPCs Table 3. Differences of georeferencing and RPCs re-projected points and GCPs Figure 6. Coordinate differences of georeferenced (left) and RPC (right) projected points; differences displayed with scale factor CONCLUSIONS Georeferencing of image data can achieve 2D position accuracy to pixel-level. The accuracy of an ortho generated image and its assigned georeferencing information is dependent on the DEM and the quality of the sensor model. This paper discusses the RPC sensor model based ortho generation with and without bias-correction. It shows that the results of georeferencing can be significantly improved by bias correction of the RPC parameters prior the ortho generation. Georeferencing via GCPs was successfully applied to ortho-rectified image data using 2D affine transformation in different coordinate systems. Georeferencing of non-rectified image data can only provide coarse results and should only be applied to local areas or areas with moderately flat terrain.

6 The orthoimage generation with bias-corrected RPCs were performed with the software package Barista and shows its applicability to produce accurately georeferenced image data. Barista is also able to handle provided georeferenced image data, as well as the necessary coordinate transformations from image to object space and vice versa. Georeferenced image data enables extraction of 2D object space coordinates from images for a broad user community with a range of applications without requiring any special photogrammetric background. This opens the use of georeferenced image data to a wider audience and assists in many geo-related applications. REFERENCES Adobe Developer Association, TIFF Revision 6.0, 1992, Baltsavias, E.P., Digital ortho-images a powerful tool for the extraction of spatial- and geo-information. ISPRS Journal of Photogrammetry & Remote Sensing, 51(1996), pp DigitalGlobe, DigitalGlobe Ortho Imagery, CO, USA. (accessed 26 Sep. 2006) ESRI, Technical Articles, CA, USA. es.articleshow&d=17489 (accessed 26 Sep. 1999) Fraser, C.S., Dial, G. and Grodecki, J., Sensor orientation via RPCs. ISPRS Journal of Photogrammetry & Remote Sensing, 60(3), pp Fraser, C.S. and Hanley, H.B., Bias compensation in rational functions for Ikonos satellite imagery. Photogrammetric Engineering & Remote Sensing, 71(8), pp Grodecki, J. and Dial, G., IKONOS geometric accuracy. Proceedings of ISPRS Joint Workshop: High-Resolution Mapping from Space, Hanover, Germany, September 2001, 8 pp. (on CD ROM). Ritter, N. and Ruth, M., The GeoTiff data interchange standard for raster geographic images. International Journal of Remote Sensing, 18(7), pp ACKNOWLEDGEMENTS The image data set of Hobart was provided by the courtesy of Space Imaging (GeoEye). The image data sets of Bhutan and Moreton Island including ground control information were provided by Digital Globe/SKM. The authors want to thank the image providers for supplying their data sets to make the analyses for this paper possible.