Вы находитесь на странице: 1из 11

Pat Mroczek, GeoFficient Strategies GISC 9216 Deliverable 3 Geometric Correction, Orthorectification and Mosaicking

12 March 2014

Geometric Correction, Orthorectification and Mosaicking

Introduction
Digital imagery contains geometric distortions that occur from variations in altitude, attitude, and velocity of the sensor platform to factors such as earth curvature, atmospheric refraction, relief displacement, and irregularities in the path of the sensors IFOV (Instantaneous Field Of View)1. The two type of distortions- systematic and random- must be corrected to obtain a spatially accurate image. Geometric correction can alleviate these random distortions through the use of Ground Control Points (GCPs). GCPs match points on non-georeferenced images with points on a georeferenced image. Three digital images overlapping an area south of Guelph, Ontario near Aberfoyle, Ontario, were corrected and orthorectified to produce two mosaic results. Mosaicking is the process of merging multiple raster images together to form one unified image. These processes provided practical experience correcting and registering digital imagery to a previously georeferenced image.

Data
Three uncorrected TIFF images were used for geometric correction: Photo_1, Photo_2 and Photo_3. The georeferenced base image covers an area southeast of Guelph, Ontario, approximately 8km by 7km. A road network file and a building shapefile were the two vector files used to compare the accuracy of the mosaics.

Methodology
Two geometric models were used to produce two sets of corrected digital imagery. The polynomial geometric model uses polynomial coefficients to map between images 2. First order polynomials were used to transform the images into a UTM Transverse Mercator projection. There are limited parameters to set when using this method. Eight GCPs were used to correct each image. These points were spread out through the entire extent of the image in order to represent the whole study area. This is critical to allow the rectification process to stretch and fit the entire image to the base map.

1 2

Remote Sensing and Image Interpretation, 6th Edition, Lillesand and Kiefer, p. 487 ERDAS Imagine 2013 Help Files 1

Pat Mroczek, GeoFficient Strategies GISC 9216 Deliverable 3 Geometric Correction, Orthorectification and Mosaicking

12 March 2014

Intersections and distinct points of interest were chosen for the GCPs. Figure 1 shows example of 8 GCPs across the study area.

Figure 1: Spread of GCPs Across Study Area (Left Unrectified Image, Right Base Image)

Once four points were selected on the uncorrected image and matched with the corresponding base image, the Root Mean Square (RMS) error was calculated. This value shows how well the model predicts the locations compared to the base image. The RMS is a mathematical function that takes into account distance between two points in the context of a group of points (Such as 8 GCPs in this case). Low RMS values indicate high accuracy. The image is then resampled based on the GCPs and rectified to the base image. Once completed for the 3 images, each image was clipped to remove the unnecessary black boundaries. This was done by creating an Area Of Interest (AOI) file and using the AOI to create a subset of the image. The three images were now ready to be mosaicked. This was completed through the MosaicPro tool by simply adding all three images and running the mosaic. The second method used the camera model to perform orthorectification. Camera Model properties were inputted into the model as per the Terms of Reference 3, based on the camera that was used for the images. This method uses a Digital Elevation Model (DEM) to remove relief displacement and uses the focal length of the camera as an input. The fiducial points on each image were identified and plotted to help in orthorectification. These points were identified as in Figure 2.

Terms Of Reference GISC 9216 2

Pat Mroczek, GeoFficient Strategies GISC 9216 Deliverable 3 Geometric Correction, Orthorectification and Mosaicking

12 March 2014

Figure 2: Fiducial Points Linked From Computer Image to Film

Eight GCPs were selected for each image in a similar process as above. The Solve function was enabled for the model to start predicting RMS errors as opposed to automatically showing up like in the previous method. Points were assessed for accuracy by the RMS error and removed if values were above 0.1, indicating lower than ideal accuracy. After all points were selected, the image was resampled for orthorectification. The three images were clipped with the use of AOIs and mosaicked using the MosaicPro tool. The two mosaic results were overlaid with the base image to compare the two methods (Appendix 1 and 2).

Polynomial Correction and Mosaic


The prediction process provided a good localization of each GCP that was entered. Localization was not exact but generated a point within the vicinity of the actual point. Figure 3 shows an example of GCP #5 localization.

Figure 3: Localization Example For Polynomial Method. Red Dot shows actual location 3

Pat Mroczek, GeoFficient Strategies GISC 9216 Deliverable 3 Geometric Correction, Orthorectification and Mosaicking

12 March 2014

The prediction process improved as more GCPs were selected. The variation in accuracy is based on the RMS errors associated with the previous points. High errors would provide a localization that is further from the actual location while low errors would provide much more reliable predictions.

The control point error is an aggregated result of the amount of error from the GCPs. When a GCP is positioned, a prediction point is made on the base image for the corresponding point. If the point is not predicted correctly, it can be adjusted and the distance from the actual point to prediction point is quantified into an RMS error. Table 1 below shows the error for each image.
Table 1: Polynomial Control Point Errors In Metres

Method Polynomial

Image 1 2 3

Control Point Error 0.0496 0.0618 0.0582

The first image had the lowest error but all three have relatively similar control point errors lying near the 0.05 to 0.06 range. This error is suitable since there are no significant discrepancies between image errors and errors are not substantially high or above 0.1.

The original pixel sizes of the TIFF images are 0.0025cm (0.000025m). The images come from the same sensor in the same conditions and at the same time, therefore they will all have the same pixel size. Conversely, the default pixel sizes for the resampling of the three images are all close to 0.50m as seen in Table 2.

Table 2: Resampling Default Pixel Sizes (m) for Polynomial Method

Image

Pixel Size 1 0.501361809 2 0.49994292 3 0.500586669

They vary slightly above and below the 0.5m threshold since post-processing has altered the images. The reason behind the slight differences is because of the different spread of GCPs for each image. The image needs to be stretched and repositioned for geometric correction which causes the change in pixel size from original to processed image.
4

Pat Mroczek, GeoFficient Strategies GISC 9216 Deliverable 3 Geometric Correction, Orthorectification and Mosaicking

12 March 2014

The quality of the mosaic turned out very well (Appendix 1). The road network file matches up with the mosaic fairly well. The buildings shapefile overlays actual buildings fairly well except for a slight shift in some areas (Figure 4).

Figure 4: Section of Polynomial Mosaic With Building Overlay

It is evident that some of the building polygons do not lie exactly on the mosaic image. The cause of this shift is inaccuracies in the geometric correction. It could have been improved by slightly more accurate and better spread out GCPs. The road network file overlaid on the mosaic shows some shifts in the actual location of roads (Figure 5). Yellow boxes indicate areas of discrepancy between mosaic and shapefile. This may have been improved with a better spread of GCPs. The area above the top yellow AOI did not contain many GCPs because of the difficulty to locate suitable points. Additional GCPs in this area would have stretched the top of the image to correspond with the road file better.

Figure 5: Road Network for Polynomial Correction 5

Pat Mroczek, GeoFficient Strategies GISC 9216 Deliverable 3 Geometric Correction, Orthorectification and Mosaicking

12 March 2014

Orthorectification using Camera Model


The prediction process for orthorectification is giving a much better prediction compared to the polynomial method. This prediction could be improved partly because of the inclusion of a DEM to reduce relief displacement. Also, the addition of fiducial points boosts the accuracy of selected GCPs.

This method provides much lower control point errors compared to the polynomial method. As mentioned above, the localization is more accurate, therefore the error will be greatly reduced. Table 2 shows a comparison of control point errors between the two methods.
Table 3: Comparsion of Control Point Errors In Metres

Image 1 2 3

Control Point Error Control Point Error Polynomial Camera 0.0496 0.0125 0.0618 0.0112 0.0582 0.0056

The errors are only a fraction of those for the polynomial method. The addition of several parameters such as the DEM, focal length, and fiducial points improve the effectiveness of the model and prediction process.

The default pixel sizes for the orthophotos are the same as the process above, around 0.5m. Table 4 shows the pixel sizes of the three images are just above 0.5m.
Table 4: Resampling Pixel Sizes (m) for Camera Method

Image

Pixel Size 1 0.501170386 2 0.500764066 3 0.500638107

Again, the difference is caused by image processing. The image is stretched and fitted to the base image based on the GCPs. Since there are different GCPs on each image, each resampling process will be slightly different. It makes sense that both methods produced similar pixel sizes since it is the same image being processed and fitted to the same base image.

Pat Mroczek, GeoFficient Strategies GISC 9216 Deliverable 3 Geometric Correction, Orthorectification and Mosaicking

12 March 2014

The quality of the orthomosaic is very accurate (Appendix 2). Features on the mosaic appear to line up very well with the base map image. All buildings on the mosaic overlap with the buildings in the shapefile with minmial shifts in some parts (Figure 6).

Figure 6: Portion of Ortho-Mosaic Showing Building Overlay

Any shifts could be improved by a better spread or a higher number of GCPs. The road network shapefile can be viewed in Figure 7. Again, the yellow boxes indicate the same AOIs as the polynomial mosaic. The improvement is evident in the bottom two AOIs but there is still an inaccuracy in the top AOI. One or two GCPs above this AOI could have produced a better orthorectification, stretched to correspond with the road shapefile. Inaccuracies in the road shapefile could also be the cause of the mosaic not lining up.

Figure 7: Road Network File Overlaid On Ortho-Mosaic 7

Pat Mroczek, GeoFficient Strategies GISC 9216 Deliverable 3 Geometric Correction, Orthorectification and Mosaicking

12 March 2014

Conclusion
Orthorectification provided a more accurate mosaic than the geometric correction method. The highest control point error for orthorectification was 0.0125m compared to 0.0618m for geometric. Orthorectification predicted GCPs on the base image at a greater accuracy than for the geometric model. The number of GCPs and the spread these GCPs are integral to the quality of rectification. It is important to keep this in mind when following through with geometric correction or orthorectification since well-placed GCPs will provide highly accurate results. Addtionally, the use of a DEM, fiducial points, and focal length contributed to the improved model of orthorectification. The geometric model only used GCPs as inputs for correction. Since the mosaic procedure was the same for both methods, the correction model causes the difference in final results. Orthorectification is the ideal model to use when the data is available.

Pat Mroczek, GeoFficient Strategies GISC 9216 Deliverable 3 Geometric Correction, Orthorectification and Mosaicking

12 March 2014

Appendix 1: Polynomial Correction

Pat Mroczek, GeoFficient Strategies GISC 9216 Deliverable 3 Geometric Correction, Orthorectification and Mosaicking

12 March 2014

Appendix 2: Orthorectification and Mosaic

10

Pat Mroczek, GeoFficient Strategies GISC 9216 Deliverable 3 Geometric Correction, Orthorectification and Mosaicking

12 March 2014

Bibliography Intergraph. (2013). ERDAS Imagine 2013 . Set Geometric Model. Niagara College GIS. (2014). Geometric Correction, Orthorectification and Mosaicking. Deliverable 3. NOTL. T. Lillesand, R. K. (2008). Remote Sensing and Image Interpretation. Hoboken: John Wiley & Sons.

11

Вам также может понравиться