Introduction. Mathematical Background Preparation using ENVI.

Transcription

1 Andrew Nordquist Investigating Automatic Registration and Mosaicking in ENVI 3 December 2007 Project Proposal for EES Remote Sensing Class Introduction. Registering two images means we want to compare data between the two in a quantitative manner. The images may be rotated, displaced or even scaled differently compared to one another, depending on the source of each image. For instance, one may want to compare Landsat data with an equivalent Hyperion image. Usually, performing a registration means manually locating many points known as ground control points (GCPs) - that are known to be the same feature between the two images. It also turns out that automatic (computer) registration of two images can be done, provided there is enough common area. It may be possible to use this same technique to mosaic multiple adjacent images together into one image. Automatic image registration of two square map portions can be done using a Fourier Transform technique. This technique is described by Xie, et al, in a paper published in the Computers and Geosciences journal in March 2003 [1]. The paper was entitled, An IDL/ENVI implementation of the FFT-based algorithm for automatic image registration. I will go over what I did in ENVI and with the IDL routines provided from the abovementioned paper. When I refer to IDL routines, I am using them verbatim from the ZIP file provided by a link at the LRSG website. I will show results from my attempts to perform automatic registration on files other than what was provided. Mathematical Background. It turns out that if you perform a fourier transform on each of two images whose difference is only a shift, there is a simple relationship between them [2]. If you take the inverse fourier transform after normalizing by their absolute values, you end up with a delta function that is located at the point in the original image space that is equal to the shift, from the origin. This multiplication of two fourier transforms is very similar to what s known as a convolution, or a correlation. This is not to be confused with the statistical covariance of two distributions. Perhaps the fourier transforms give you an analog in a two-dimensional space to the rather one-dimensional distributions of numbers. Preparation using ENVI. I explored ENVI's capability for doing what was termed Automatic Registration. I compared registration techniques with ENVI and with the IDL routine provided. One example was a simple shift in the pixels. Both methods picked up the correct amount of shift. Another example was a simple scale and rotate. Again, both methods were successful in finding the amount of shift and rotation that was introduced. By both methods, I refer to 1) the Automatic Registration capability of ENVI, plus 2) the IDL routines. ENVI s Automatic Registration feature is located under the Map menu. The automatic part involves automatically generating GCPs so that the images can 1

2 be registered in more or less the old style. I will get into the details of the ENVI procedure shortly. Another example provided was a little more interesting. It was a set of Landsat images of nearly the same place at different times. One image was shifted from the reference; another was scaled and rotated. Again both methods gave expected results. By this point in time, I have repeated the work done in the Xie paper, but did parallel work in ENVI. The trouble was, ENVI work took a lot more time than just running the IDL routines, because of having to deal with GCPs. To proceed with Automatic Registration, select Map Registration Automatic Registration: Image to Image from the ENVI menu. There are other options that allow for working more directly with GCPs. Then it s time to select base and warp images. The base image is the unaltered reference file; the warp image is the one that will be altered to register with the base image. Next, some Automatic Registration parameters are selected. There s a choice for matching method: area-based or feature-based, with different parameters for each. I ve had trouble with the area-based matching method on a more complex image. I ve gotten better GCPs with the feature-based matching method. However, with a simpler region (i.e. the shifted image provided from the paper), I got excellent GCPs from using the area-based matching method. There s an option for examining the tie points before warping the image; I think it s always a good idea to examine those points. I ve seen some poor automatic choices with the cases that I ve run. If feature-based matching is chosen, some iteration goes on automatically. The program switches between building scale space images from image pyramid and finding feature points from difference of Gaussian, going through each procedure twice. Then it matches feature points and presents you with a GCP Selection window and Image-to- Image GCP list. After examining and adjusting the points, they can be saved. I found that if I was able to reduce the error that was displayed in the GCP table to less than 1 meter, I got good results on warping. Under the Options menu of the GCP Selection window is the provision to Warp Displayed Band. Up pops a Registration Parameters window. There are choices for method between RST, Polynomial (with degree), and (Delaunay) Triangulation. There are also some Resampling choices: Nearest Neighbor, Bilinear, or Cubic Convolution. I only had a chance to use the Polynomial method, up to degree 3 (depending on the number of GCPs), and the Bilinear or Nearest Neighbor resampling. Once the choices are made, the warped image becomes available. I ve seen some very good and some very bad warped images in the time that I ve worked with the procedure. If care is taken to get good GCPs, the warped image provides an excellent match to the base image, and a mosaic can be done (see below). Once the warped image comes away from the region where the two images overlap, it can warp quite noticeably. 2

3 Sometimes I ve seen it, and sometimes I haven t. It seems to depend on many things, not the least of which is GCP selection and registration and resampling methods. The following image shows the mosaic I made from two of the images provided from the Xie Paper. There is some distortion in the warped image where it doesn t overlap with the base image; I think this is an average example with distortion that could be improved with more work on the GCPs and the registration methods. So there it is - the Automatic Registration procedure in ENVI. It s not so automatic; then again I would not want to select 25 GCPs manually. I think that would take longer. I then proceeded to find some images on the Internet that I could use the IDL routine on to find out if I could extract information from the process. I thought storm or hurricane weather patterns of the infrared or water vapor type would provide good input data. I found and downloaded GIF files from the NOAA website. They have an archive database going back up to 3 weeks on hourly weather images for the West Coast, Alaska, and Hawaii. I picked up one set of each, covering 27 hours of weather for the three areas, with four hours difference between the images. The images for Alaska are on the following page and Hawaii images are shown on the page after that. The set of Alaska images I think looks like a difficult thing to register; while the Hawaii images look like they would do a better job. 3

4 4

5 5

6 I thought a much better image to register would be a hurricane, so I was able to find GIF images of Katrina and Rita. The Katrina images are shown two pages following this one. Surely those two images would show some kind of shift or rotation! Results. I did not get any shift or rotation results from any of these images. I looked at the statistics and I think I found the problem. The histograms are pretty choppy; not all DN values are represented in these images. I had to convert the GIF files to TIF before I could read them into ENVI; and I tried JPEG, Bitmaps, PICT, and one or two other types, but the histograms remained the same. I stuck with TIF format. I always got three bands of data, with each band usually identical to the other two. The following page shows representative histograms for the data I obtained. I tried stretching the Alaska data, but the outcome did not change. I was able to bring the GIF file into ENVI with several different types of bitmaps: monochrome, 16-bit, 24-bit, and 256-bit. Each one gave me a different histogram, with more bits generally giving me more 0 s in the DN values between nonzero points. That is, there were more unused DN values the more bits worth of bitmap I tried. I think the solution rides within that clue: I think I need to read the GIF files into ENVI in a different way. Or, I need to process the GIF file in a different or special way. I m interested in pursuing this, to see if I can resolve the situation. Bibliography. [1] Xie H, Hicks N, Keller GR, Huang H, and Kreinovich V. An IDL/ENVI implementation of the FFT-based algorithm for automatic image registration. Computers and Geosciences 2003; 29: [2] Araiza R, Xie H, Starks SA, and Kreinovich V. Automatic referencing of multispectral images. Proc IEEE Southwest Symposium on Image Analysis and Interpretation 2002; pages

7 Histograms for the weather plots. Top, Alaska; middle, Hawaii; bottom, Katrina. 7

8 Katrina images from GIF files. These two images have the same scale and geographical placement, as indicated by the outline of the FL, LA, and TX coast shown in white near the top. It should be an ideal pair of images to attempt to register to each other. 8