Hi,

Place More Ground Control Points (GCP), While doing georectification.

Introduce more GCPs, wherever you find more mismatches.

The distribution of GCPs should cover all over the image.

Reduce your RMSE value as much as possible.

Select good Ground control points, because it should be identified in both the images.

Use ERDAS Imagine ---> the option Geometric Correction is a good software to do georectification process.

Thanks Ramesh for suggestion. I am followed all these steps using ERDAS and ArcMAp, but still not satisfactory results. Actually my study area is mountainous. Both images are already ortho-rectified.

Then you can improve your quality with DGPS Survey.

However you have very high resolution images, take some reference points from the Google Earth Image and use for your georectification process.

I aggre with Veerappan Ramesh.

You can use ENVI Ortorectification Module

I was facing similar problem a couple of years before, I fixed them by:

- collecting GCPs from all parts of the image (random and even GCPs distribution)

- taking stable GCPs, road cross, building corners,

- more CGPs (with less error)

NOTE: area under investigation was NOT mountainous.

With high resolution image data height differences can cause height displacement of the order of the sensor resolution. If the images were not observed from exactly the same point in space, then they can have different displacements, which could cause mis-registration errors. Although you do say they have been orthorectified already - it might be worthwhile checking the process by which this was done and the digital terrain model or contours which was used to do the orthorctification. Could you also check that you did a first order registration FIRSTand have checked the residual errors. Too often a 2nd order registration is chosen and any errors in control point selection are not picked up.

Since the topography is complicated, one shot correction will not be suitable. I suggest to divide images into flat and non-flat areas and georectify each of them separately.

Sorry for being self-referential, but may I propose you to use the software I created?

It is ABSOLUTELY FREE (it just asks an email to keep you informed about updates) and new case-studies are always welcome. It makes automatic matching (scale, rotation and even color invariant) and georeferencing in few seconds and I am also working on a new version that improves the successful matching. In the worst case... you lost five minute to try it out :)

I will be happy to help with it.

You can download your copy of AutoGR-Toolkit at

http://www.ims.forth.gr/AutoGR

Please let me know if that helps.

Cheers,

Gianluca

Hi,

I had similar problem - 2 VHR images with mountains wchich didn't fit.

Probably during ortorectification they used too pure DTM.

the best option I found was in ArcGIS:

Toolbars-Georeferencing and for transformation choose "spline" - if fits image to image localy

spline option is "invisible" if you don't have enough points (at least 10)

you should set many regularly distributed tiepoints to fit images well

In order to use these images properly you must make an orthorectification, using D-GPS (Differential GPS) points obtained during field survey. These points must be equally distributed through the entire image, so that get a good quality work.

Hi

In high resolution images it is necessary to pay attention 3 important points

1.number of Ground Control Points

2. distribution of these points

3. transformation method

i am totally disagree with dear Michal because spline method will destroy your image because of its adjustment .

2nd or 3rd order polynomial method are suitable for your target.

if you want very accurate georeferencing i supposed to use Ortho engine tools in Geomatica software which let you to have 3d rectification. and there you can decrease you error by DEM data.

best regards

as far as I see the problem - there is no time for colecting Ground Control Points and the only data Hammad has are two ortorectified images which do not fit each other.

I totally agree that spline transformation may destroy the image if you have only few points - it would be a dissaster but if you use a lot of points - like 100 and they are distributed regularly (very important) the result is very similar to 2nd and 3rd order polymonial transformanion.

the advantage of spline method is that if polymonial transformation are not able to fit 2 images perfectly, with spline you can force part of the image to make it fit better

to do this firstly try 2nd and 3rd order polymonial transformation (with regularly distributed points!) and than swich to spline to see the diffrence and judge iwhich transformation gives better results

I've done lots of test about that, comparing diffrent methods for "bedly rectified" images and - it was supprise to me too that spline was the best for this task

@Lachezar Filchev

I didn't know about the specific software you mention, I will take a look at them and try a comparison.

Thanks for sharing.

Mohamed Eleiche: I can't separate flat are mountainous areas. because its a watershed.

Michal Krupinski : "spline" destroyed and completely reshaped my image. I tried 2nd and 3rd order polynomial transformation using ArcMap and ERDAS Imagine but still not getting the satisfactory results.

Thanks for Gianluca Cantoro for the suggestion of AutoGRT, I will try and let you know the results

Lachezar Filchev : I will explore Spring and Orfeo Toolbox.

The point in the question is that, in line of principle, the images should not be altered, but only the relative georeferencing. If you are confident in the DTM and georeferencing of the first image you should compute the registration error by a pattern matching procedure and then correct it. SIFT and SURF algorithms could be suitable in your case, since these algorithms allow to achieve a sub-pixel accuracy in many cases. I suggest you to use the SW of Gianluca Cantoro or similar.

In case of VHR images, very accurate independent reference source such as DGPS would be a good option to rectify both images, instead of going for one image as a master reference and the other as slave. In case of elevated structures, it might cause a problem due to viewing angle difference. Thats what I can say.

Shridhar Jawak : I agree with you, 2009 image is orthorectified using the RPC file and high resolution DEM by taking care DGPS survey that time. Based on the 2009 image I developed number of layers (most important Community Forest User Groups boundaries and delineated tree crowns), so I am bound to take it as a reference. While 2012 image is sensor based orthorectified (GeoProfessional).

I've used the spline algorithm (PCI) for airborne scanner data to vertically register several image data sets. You do have to use many points, finding the correct number of points is important. I haven't tried the spline in ARCGIS. The spline algorithm does a local correction around each point, this is very different from using a polynomial that distributes the correction throughout the image. The number of points is critical, too many or too few can warp your data around the point that is either misplaced or not needed (it looks kind of neat, but not what you want). I generally just add points that are evenly distributed throughout the image. In PCI the spline also uses a DEM. I would use the highest resolution DEM you can find.

If you are using the data for temporal analysis I wouldn't recommend independent registration of your two scenes. You will find that they are off (if only slightly in certain areas of the images), I've seen this with high resolution (1-4 m res) as well as moderate resolution (10-60 m res) data sets. If your data sets are currently independently registered a simple warping (1st or 2nd order polynomial) between the scenes may be good enough. This worked well for me when vertically registering moderate resolution data sets (ALI, Landsat, and AWiFs). I had 6 dates of the 3 types of sensors, I used the highest res data set as my reference, I then warped each data set to the reference using the same points each time. This will help cut down on possible introduced errors from the registration.

Hammad Gilani: I have suffered from the same problem in case of VHR images of urban area with flyover bridges mis-matched. I could resolve it by dividing images on the basis of height difference. And i could rectify them separately and mosaicked at the end. I could improve the results. Bit Its not what we can call it good for urban change detection studies. Just shared my experience.

Gianluca Cantoro : I have put my images into your tool but unfortunately no output even after waiting 48hrs processing (each image size is10GB).

Processing on Sample Images: First I converted my images into tiff format using your tool. Second using AotoGR Sift tried to rectify but no error/output. Please cheek and let me know, sample images are enclosed in given below link;

https://www.dropbox.com/sh/hy5gs16e5deyzmx/KKMMwnnA39

@ Hammad Gilani:

sorry to hear about it. I will need some more information to help you.

How does the converted (tif) image look like? The convertion in my software is done with gdal, and sometimes it doesn't handle the colos of the images in the proper way. Can you send me also a sample of the same area for the other satellite image? I can try to process them again and see what happens. I can also try and process them with the new version I am working on to see if things are better.

Otherwise, I will need all the details of your procedure to reproduce your condition and understand the error.

Thanks,

Gianluca

OK, sorry, I looked into the dropbox folder before the download was completed.

I got the two images and the problem has to do with the color stretching and standard deviation: both parameters are lost via gdal so what you put into autogr-sift was just a "dusty" image. I tried with a converted one and I am uploading the preview of the matchings (almost 700 points in 1-2 minutes).

Ok. Then what I have to do to solve this problem? Please guide me, so that I can run smoothly.

OK, so. First thing is to export the images from Erdas IMG format to 8bit TIFF.

To do so, you right click on each image in arcgis and choose data>export data; make sure that Use Renderer and Force RGB are checked and save them as TIFF. Now you can load them into autogr-toolkit and process them. If the previous test was with a screenshot and it found 700 matching points, with the full size input I got 1200 matching. Keep in mind that this high number of matching point may crash arcmap...but this is not my fault :)

Let me know if that works. I will now put the matching point into the shared folder for your consideration (be carefull not to opoen if with the normal notepad... unless you want to wait hours to be loaded).

Good luck,

Gianluca

Thanks, now it work at my end but I am confuse and surprise, it lost the project and produce number of files. And most important, still I can see spatial difference in output PNG files. Anyway, Thanks a lot for providing this tool, need more refinement.

Please read the help file and the readme.txt to understand the meaning of each produced output. Spatial differences will always be visible unless to align pixel by pixel... And still you can add reference point to the ones produced by the software. Definitely AutoGR-Toolkit needs some refinement (also because I am not a programmer, I am archaeologist). By the way, feedbacks are always welcome.

Thanks to all of you who will drop me a line about how to improve it.

Cheers,

G.

AFAIK most of the software in use for this application rquires that you have control points (i.e. zero-dimensional features). I have developed recently an algorithm to transform one dataset (either raster or vector) over another by matching routes, fences or any one-dimensional feature which could be identified in both datasets. Since you claim that the images are already ortorectified, hopefully all DEM-related displacements have essentially been considered, so it is fit to apply other methods. The procedure has been described in a recent paper, only available online at http://www.tandfonline.com/doi/abs/10.1080/13658816.2012.677537

I could supply source code in matlab if someone wants to make a try.

Regards

Carlos

Thanks. I downloaded your paper, first let me read and then come back to you for code.

Meanwhile if you can explain "Is this code can handle VHRS data (~size 10GB each image)? In my case how much you are confident that your code can match pixel to pixel?

To me earlier tool and your one look like more research point of view nice but operationally I have question marks. Thanks and looking forward to hear you.

You posed some different questions.

a) How much confident? As usual, it depends on your dataset. Do you have plenty of control points, or none? With this particular algorithm: do you have roads or any other linear features which appears in both images? How many, and where? What about the displacement errors due to topography? Hard to be confident ...

b) The algorithm can deal with images of any size: it is just a matter of coding. The present code is limited to its working environment (matlab and probably clones) so it is likely that you will be in trouble. However, you might want to consider some workaround, making part of the task in the Matlab environment and the final, heavy work directly with your large images in another environment at your wish. This might require some minor coding from your side. I can explain this a little bit if you want to, but I feel it might go out of topic for this thread.

c) Yes, it is still research code. I mention it in this thread because apparently there have been no success neither alternatives with out-of-the-box commercial code, or even with some other research code.

But if you have no polylines (roads, fences, etc.) to match in both images, the code will be of no use.

Carlos,

I have images that need to be aligned based on another image that is properly geo-referenced. Actually, these images are based on agricultural fields and have very high resolution (i.e., less than a meter). The plots within the field are distinct in the form of lines. Do you think your program can handle this situation. Can you please supply your source code in Matlab so that I can test with my images. Thanks

Hi, if you didn't try it already, you can download the new version of my software for automatic image georeferencing

www.ims.forth.gr/autogr

From our first contact a lot of work has been done, so hopefully this new version can manage to do things that the version of 2013 couldn't.

I'll be happy to hear from you whether it worked or not.

Good luck ;)

Thanks Gianluca, 

I tested this software but it failed to operate with my images.