Forum

Author Topic: Historic Aerial Imagery Advise  (Read 6282 times)

Rossta

  • Newbie
  • *
  • Posts: 2
    • View Profile
Historic Aerial Imagery Advise
« on: February 03, 2015, 05:33:38 AM »
Hello,

I am fairly new to using this fantastic piece of software. My end goal is to come to come up with an automated procedure to orthomosaic many, many scanned historic images. I have been reading about working with digitized imagery on these forums and I was hoping someone could take a look at my workflow/data and let me know if I am missing something.

A little about my data:
The images are from 1954. They were hastily clipped/georeferenced so they don't align well and the warping of the image means none are the exact same height/width (which means photoscan treats them as a separate cameras). The fiducial marks are clipped out and someone drew/wrote on the original imagery which I think may be affecting alignment. All photos have a focal length of 152.3mm and were flown with 60% stereolap at 12,500 feet AGL. I have samples available here if anyone would like to take a look: ftp://ftp.csc.noaa.gov/temp/psc/HAI

A little about my workflow:
1. I generate an image centerpoints text file in GIS software according to the format in the PSPro Manual to assign reference coordinates
2. Load images into photoscan (1 chunk, x number of cameras)
3. Mask as much of the writing as possible using the magic wand tool
4. Align images using high/generic as the settings
5. Build a dense point cloud using the aggressive depth filtering
6. Build a mesh
7. Load in the reference coordinate file to get WGS84 UTM 5N coordinates
8. Export as an orthomosaic

I have had some success, but I am still getting tons of artifacts in my orthos (jagged pieces along shorelines/valleys/etc.) There is a sample ortho export in the ftp link above. Any advice would be greatly appreciated!

Thanks in advance!

StevenF

  • Newbie
  • *
  • Posts: 39
    • View Profile
Re: Historic Aerial Imagery Advise
« Reply #1 on: February 03, 2015, 10:25:36 AM »
Hi Rossta,
I've also been working many (hundreds to thousands) of historic aerial images that have some of the same problems as your images (e.g. different sizes, no fiducials, writing, etc.). I can make a few suggestions, but I'm also still trying to figure out the best approach.

With regard to the artifacts in your ortho, I'd recommend looking at your mesh and point cloud for sharp breaks and discontinuities that could be manually edited to be smoother. Also, look for large gaps in the point cloud. A common problem I've had with poor quality imagery is a lack of completeness in the point cloud particularly when using High or Ultra quality settings. I tend to get most complete coverage with Low quality and Aggressive filtering in dense cloud generation, but these settings will produce lower point densities overall and probably have lower accuracy.

If the mesh and point cloud look ok then you may want to consider mosaicing the individual ortho frames in another program that allows seamline editing. I've seen these artifacts before even in areas where the mesh is smooth, and I think it has to do with low overlap and the way PhotoScan selects an image to map on the mesh.  You can export individual ortho's by disabling every camera but one before exporting. If you have many photos you'll want to implement this in the python api. Some other GIS and Remote Sensing programs like ArcGIS are capable of mosaicing images with seamlines that can be edited.

Here are a few other suggestions for changes to your workflow:
1. Use ground control points: Instead of estimating the coordinates of the image center points for referencing, collect ground control points (gcp's) of identifiable locations in the photos. You can get the most accurate gcp's with a survey grade GPS, but getting coordinates from existing orthomosaics (like Google Earth) can yield accuracies that might be sufficient depending upon your purpose. Your existing ortho was off by around 500m in some places, but you commented on the artifacts instead so I'm guessing positional accuracy isn't that important to you.

2. Optimize alignment: Once your ground control errors are reasonable you can often improve accuracy by optimizing alignment which will use the gcp's and tie points (i.e. sparse cloud) to refine the individual camera locations and camera calibration parameters. Without 'optimizing' I think ground control is merely used to translate, rotate and scale the whole set of cameras together.

3. Alignment via referencing: If you ARE going to collect image centers for referencing then you might as well use them during alignment. Load the reference data for each image before alignment and use the Reference pair preselection setting. It should significantly speed up alignment when working with a large number of images since PhotoScan will know to only look for tie points between adjacent photos. You may find that using reference data during alignment is necessary when you step up to big projects with poor quality images.

4. Camera Calibration: If you know camera calibration info like the focal length, pixel size, or a rough guess of the principle point location then enter it as initial information in the Camera Calibration dialog.

I'm also still trying to figure out the best approach for big projects with historic imagery so please share your results if you find a better approach, specific settings, image prep, etc. Good luck.
« Last Edit: February 03, 2015, 10:30:32 AM by StevenF »

bigben

  • Sr. Member
  • ****
  • Posts: 406
    • View Profile
Re: Historic Aerial Imagery Advise
« Reply #2 on: February 03, 2015, 04:21:50 PM »
Do you really need a 3D reconstruction to get an adequate ortho image? Have you tried just a traditional mosaic using georeferencing/rectification using something like QGIS or GlobalMapper?

Whilst not quite the same source data (photos were already composited and annotated as maps) I converted some of these: http://www.lib.unimelb.edu.au/collections/maps/historical/1945melb/  into this set of georeferenced images http://files.digitisation.unimelb.edu.au/kml/1945melb/l_sheets/doc.kml

aggieair

  • Jr. Member
  • **
  • Posts: 93
    • View Profile
Re: Historic Aerial Imagery Advise
« Reply #3 on: February 03, 2015, 11:24:55 PM »
I am commenting to follow because I am about to start a very similar project.

I also found this post about historical imagery too:

http://www.agisoft.com/forum/index.php?topic=130.msg2764#msg2764

Rossta

  • Newbie
  • *
  • Posts: 2
    • View Profile
Re: Historic Aerial Imagery Advise
« Reply #4 on: February 25, 2015, 04:22:32 AM »
Hey all,

Sorry to open the thread and then bail, but work picked up and this got pushed off the plate.

Anyways, I have incorporated some of the tips you all shared and I picked up from other posts on this forum (http://www.agisoft.com/forum/index.php?topic=130.msg2764#msg2764). I can say I have much better products now. Most importantly was clipping all images to the same height/width and making sure they are all oriented the same direction. To do this I abandoned my pre-georeferenced images and grabbed the raw tifs (which also have fiducials!). This has made for a much better product.

StevenF, I appreciate the advice to address discontinuities in the dense cloud. This is much easier than tweaking the entire workflow. Since I don't know camera calibration info I have been working off of the defaults which has been just fine. I also don't have ground control points (and accuracy is of secondary concern since these are just being used for visual analysis) so I have stuck with centerpoints to get rough georeferencing. I then only need to throw together 4-6 tie points in a GIS to fine tune the referencing the data. One strange issue is that I cannot use the referencing for alignment without breaking the software. For some reason Photoscan hates my centerpoints. I have to introduce the centerpoints as the very last step before exporting my orthos.

Bigben, you are correct that I am hunting flys with an elephant rifle here. However the attraction of Photoscan is I can automate large portions of the workflow while at the same time correcting the poor referencing. I am sitting on about 14,000 pictures covering dozens of islands and year ranges (and I don't have many resources to leverage). Image quality, pixel size, sensor type, etc varies wildly so Photoscan is much more attractive then a GIS-based approach that requires hand referencing. I really like your site though, once the data is ready I hope to serve it up in similar fashion!

I will keep posting any breakthroughs I have and links to completed data. I hope to write out my workflow when all is said and done.

Thanks!

bigben

  • Sr. Member
  • ****
  • Posts: 406
    • View Profile
Re: Historic Aerial Imagery Advise
« Reply #5 on: February 25, 2015, 08:37:36 AM »
I can certainly appreciate the desire for automation.  A mix of the techniques might also help.  Photoscan may leave holes in some areas where it has trouble with image detail, but it may be practical to take the Photoscan the georeferenced output image and then patch problem areas using traditional GIS georeferencing.  It will be interesting to see what you come up with.