Assignment 9: Georeferencing and Mosaicking

Georeferencing and Mosaicking conducted between: April 8th and April 15th, 2013.

Introduction

This week's blog post will discuss the georeferencing and mosaicking process that took place after obtaining aerial imagery from a balloon mapping rig.  Explanation and discussion of the balloon mapping process will be posted in a blog next week.

Through the balloon mapping process we obtained nearly 5000 images of the University of Wisconsin - Eau Claire campus.  Very few of these images were selected for the mosaicking process, as some of the images were blurry or excessively similar to the other images, as photos were taken every second.  The campus was broken into six areas, one for each group in the class, with three students in each group.  It was then up to each group to georeference images in that area and create a mosaic.  Each group would then place their mosaic in a geodatabase for everyone to access.

Georeferencing

Before beginning the georeferencing process, there are a couple bits of information that are necessary to understand.  First, is the RMS error. The error, as described by ESRI in the ArcGIS Help file, is the difference between where the point ended up as opposed to the actual location that was specified.  The total RMS error describes the consistency of the transformation between control points, with greater consistency as you approach 0.  This does not mean that a georeferenced image with an RMS error of 0 is perfect, as the accuracy is dependent on the control points, and, therefore, if poor control points were selected then accuracy may still be poor.

Another important factor in georeferencing is the use of transformations.  A transformation allows a raster dataset to permanently match the map coordinates of the target data, and there are multiple options to chose from.  My group chose to use a second-order polynomial transformation, as this seemed to provide the most accurate image.  A polynomial transformation is built on control points and a least-squares fitting (LSF) algorithm.  The LSF algorithm intends to derive a general formula that can be applied to all points, which often results in slight movement of the control points.  A second-order polynomial transformation requires at least 6 control points to use.

The best way to georeference images would be to have ground control points taken with an accurate GPS.  A few members of our class did this while everyone else was with the balloon.  However, at the time of data collection the group was unaware that the balloon would be going to upper campus, and so ground control points (GCP) were not collected there, which is the area my group was designated for georeferencing and mosaicking. (In the coming weeks we may use a Total Station to obtain ground control points on campus.)

For the time being I will be using a reference image and using objects in the image, such as sidewalks, pavement, and intersections, as control points.  In Figure 1 you can see some of the points I used to georeference the image.  As can be seen, buildings would be difficult control points to use, since the angle of the camera greatly affects how much of a building you can see.  This makes the mosaicking process particularly difficult at times when you are trying to put images together that have varying angles of the same building.

The georeferencing itself is rather simple.  Merely click the add control points button on the Georeferencing toolbar, then click a spot on the map where a control is desire, and then click that same spot on the reference image.  Each control point made shows up in the Link window, showing the RMS error for each individual control point created.  Once all of the desired control points are created and the image is satisfactory, the information can be saved with the image by clicking on georeferencing and selecting Update Transformation.  Alternatively, the image can be rectified, which creates a new image and permanently applies the transformation to it, which is what I did.

My first time going through my chosen images and georeferencing did not go over quite so well.  For some particular reason I was not thinking clearly and I used the tops of buildings in the balloon image and referenced them to the tops of the buildings in the reference image.  As is likely obvious to those reading this, that is a very poor method to use.  This may be fine if the image is from directly above the building, but in our images only Hilltop (the building that goes over the road) could be used for this.

Figure 1: This is the first image I georeferenced, corrected and not using the tops of buildings for control points.
The red circle on the image shows where the Total RMS Error is located.  The red line delineates
the transformation method used.  All of the control points can be seen in the image, represented
by the green and red crosshairs with a number above them.  

As shown in Figure 1, I used 18 control points for the image, which is three times the number required for a second order polynomial transformation.  Though only six are required, earlier on I managed to obtain a total RMS error of around .5 using only six control points, but the image was way off.  I found that my best images had between 12 and 20 control points, and I would recommend at least 12.

Figure 2: The second image I georeferenced.  Again the RMS is circled and the transformation is
underlined.  The image is really far off on the top right, but that was intentional.  For this image I
focused the accuracy at the bottom of the image, as this one will be underneath when mosaicked. 

I was in a hurry when doing the mosaicking and forgot to take screenshots of the other four images used in the mosaic, but the process was the same.  I tried to ensure that each image had at least 50-60% overlap, and each one focused (was most accurate) on a different area.



Mosaicking

The mosaicking procedure was relatively simple as well, yet somehow managed to turn into a pain, as there seemed to be issues, either with the server or with my ArcMap session (or perhaps me).  To mosaic the images, I simply ran the Mosaic to New Raster tool.

Figure 3: These are the settings I used when running the Mosaic To New Raster tool.

Yet again I failed to take screenshots, this time of the errors/issues that occurred.  Hopefully I will learn from this and remember that whenever I encounter an error message or the result is not what is expected, I will take a screenshot for proof, discussion, and potentially determining the reason for the error.

The mosaic of my images can be seen in Figure ?. There are several discrepancies with my mosaic.  One is that the rope can be seen in multiple images.  Given the number of images we obtained, it is definitely possible to use enough images to not have the rope in the image; however, it would take me a considerable amount of time, of which I have been unable to find.  Hopefully in the near future I will be able to return to the images and create another mosaic without the rope.

Figure 4: My mosaic based off of four images.

While I was mostly covering the road up the hill, Horan Hall, and Hilltop, my group members were focused on the other areas of Upper Campus.  Once each of us finished our individual mosaics, we then mosaicked all of ours together, to be placed in an open geodatabase for the whole class to access.

Figure 5: Our group mosaic of Upper Campus.

Conclusion

While the process was interesting at first, it grew tedious towards the end.  This may have been partly due to the fact that I realized the mosaic would not (and did not) turn out as great as I had hoped.  Given the amount of imagery I think it could certainly be done much better, and hopefully sometime in the near future I will have the time to come back and try again.  It also would have been more enjoyable had there not been repeated issues with the creation of the mosaic.  As is likely apparent after reading this post, this particular assignment was challenging in spots, and I hope to be able to come back to these images and turn them into a better mosaic.