Historical Aerial georectification and mosaic optmial workflow assistance

[williv]

Hello everyone,

I am working on a large project where I am attempting to georeference and mosaic the 1937-1938 FSA  historical aerials for several counties. The historic aerials were scanned in using an EPSON Expression 11000XL scanner at a resolution of 1800 dpi and were all saved as .tif files which have varying dimensions. So far I have followed the basic workflow and looked at various forums, but I am at the point where I'm not sure if I am going in the right direction or computing the project in the most optimal manner. My current  goal is just to align photos in three transects without any distortions and then properly align those chunks. Once I have that completed I can implement the workflow for the rest of the counties piece by piece and put them together. Unfortunately, I continue running into problems such as the output of aligning photos in a single chunk sometimes have a distorted curve instead of a flat plain, and generally only one chunk becomes Referenced which makes it so the others just end up on top of one another if I attempt to merge them.

So my questions are: Are there any suggestions for a workflow that will optimize this process? Why do some transects alignment output have the distortion where the tie points cur up instead of producing a flat surface? Why does only one chunk become Referenced and creating a dense point cloud, mesh (3D model), and texture help this process? Any assistance is greatly appreciated.

Sincerely,

Williv2

[photogrammetrix]

Hi Williv2,

just a few questions to better understand what you are working on:
- FSA stands for Farm Service Agency (FSA) of the U.S. Department of Agriculture (USDA)?
- are the aerial images that you have scanned paper-prints or the original film-roll (positive/negative)?
- was it a complete scanning, including all markings that are displayed on the image frames (fiducials, focal length, time, altitude, image counter, clock etc.?

I am asking this, because the most important thing is, that the scans have to preserve the original parameters of the inner orientation of the camera. Paperprints or even the original film-rolls may have changed their original dimensions slightly, depending under what environment conditions the were stored over all the years (Temperature / humidity).

Furthermore the scanner may not be suited best for this task, although it is a more high-end one. There are special photogrammetric scanners, especially designed for photogrammetrically "correct" scans of aerial images that do preserve the inner orientation of the camera. Please look here:

http://e-collection.library.ethz.ch/eserv/eth:25230/eth-25230-01.pdf

The problem with standard flatbed scanners often is, that they introduce additional distortions to the imagery, e.g. because of a non-linear movement of the scanning head.

You can try to eliminate or minimze that "bad influence" , if the fiducials are imaged on your scans and the original film-format and focal-length is known. Use the fiducials as control points and perform a geometric correction of your scans using standard image processing software.

kind regards

[photogrammetrix]

Hi,

I forgot one important aspect:

The images must have sufficient overlap to do the modelling in PS. The standard overlap for classical sterophotogrammetry is around 60% in flight-direction with around 30% sidelap from flight-path to flight-path.

Myy experience is, that this is not sufficient for PS to do a proper 3D-modelling. Mosaicking may work with acceptable results for historical images, when you can find a sufficient number of ground control points.

kind regards

[nadar]

I confirm what Photogrammetrix is explaining: Photoscan is not suited to your goals.

I would rather use a georeferenced reference canevas, such as a sattelite image and individually warp each historical photo to match some common reference points found in the satellite image (cross roads, buildings, etc.) using a non parametrical transform.
The "georeferencing" module of ArcGIS 10 is very efficient for this. When all photos are corrected, you can merge them in a single mosaic. To achieve this, I'm using  Avenza's Geographic Imager in Photoshop, but there are plenty other solutions.
Good luck

[photogrammetrix]

Hi Williv,

me again   

may be this is helpful for you:

http://sourceforge.net/projects/e-foto/

http://www.uni-koeln.de/~al001/airdown.html

kind regards

[Kiesel]

Not sure if this is working with historical images but you can try AutoGR-Toolkit by Gianluca Cantoro also (even in combination with Agisoft PhotoScan):

http://www.ims.forth.gr/index_main.php?c=90&l=e&d=7

kind regards

Karsten

[williv]

Thank you for the quick replies. I appreciate all of the input. I have attached one of the photos as an example of what I am working with. The photos were scanned in from the original positive photos along with dates, serial numbers, and fiducial markings. As for focal length, altitude, and scale we have done our best to extrapolate accurate measurements based off of the documents provided from Farm Service Agency (FSA) of the U.S. Dept. of Agriculture. I was not part of the scanning process, but I can provide that information as soon as my college arrives to the office.

All great points and I have considered using the traditional "georeferencing" module found in ArcGIS 10, but there are just too many photos to go in and manually digitize every photo. For example, I have been working on only 1 out of the 15 counties for the project and that county alone has 842 aerial photographs that need to be mosaicked together. Will the methods you suggested combat this problem?

Very Respectfully,

Williv2

[brownish]

Hi everyone,

I am working on the project with Williv2 and can provide a bit more information on the scanning process.

The photos were paper-prints from the Farm Service Agency. We scanned the entirety of the photographs, including fiducial marks. I don't recall what types of additional marks were present on the borders. We did request the roll information from the National Archives and we received some documentation on the brands of cameras and lens used. The specific flight altitudes were not recorded but it seems we can estimate that using the photograph's scale and the focal length. We've been trying to figure out more about the cameras but we've been mostly unsuccessful. Williv2 can provide the camera information we have if anyone could help out.

Thanks much,
Andy

[photogrammetrix]

Dear Williv, Dear Andy,

I understand that the great number of scanned images from different cameras will make a manual photo-by-photo approach very time consuming and ... boring.

As Nadar mentioned already, Agisoft PS is not well suited for your data, because it is made to process photographic images from which the software is able to calculate / estimate the inner orientation of the camera, utilizing algorithms for camera-self calibration following the concepts of Brown.

Unfortunately the inner orientation will in most cases get lost during the scan-process for reasons I mentioned already. And perhaps there are some more, e.g. the photo was not exactly aligned with the scanner frame etc. etc..

Some time ago I carried out several test-processings with scanned paperprints of WW2 recon images from German Luftwaffe. The drawback with these images was, that the bottom fiducial had been cut off and each scan had another dimension (pixel rows/ columns). On the other hand,  I was lucky, because the type of camera, film-dimensions and focallength was known, so that I was able to do a kind of geometric "reconstruction" for each single image. After that procedure some images with sufficient overlap could be aligned, but there were always images that did not.

When you do not need necessarily "orthorectified" images with very precise positioning in your target coordinate system , trying out some kind of image-stitching software my help you out.

But you have to keep in mind, that these kind of software will not necessarily take photgraphic-image distortion into account, which is still inherent in the scans, although it can not be parameterized any more.

Some time ago I tried out Regeemy, which is an automated image stitching package developed by INPE / Brasil, which i found quite good for Satellite images. Please look here:

http://wiki.dpi.inpe.br/doku.php?id=wiki:regeemy

Good luck for your project.
kind regards

[photogrammetrix]

Hi Williv, Hi Andy,


may be this is helpful for you:

http://www.isprs.org/proceedings/XXXVIII/1_4_7-W5/paper/REDWEIK-119.pdf

kind regards

[bigben]

I ran a very small pilot project looking at a couple of scanning options for large volumes of negatives.  I only had 4 negatives, but I'd also received a set of 9 images that had been previously scanned, as well as individually georeferenced versions of these. Some early notes here: http://www.agisoft.com/forum/index.php?topic=3790.msg19777#msg19777 and a model created from these https://skfb.ly/F9FE

In the end I used an image template of the fiducial marks to manually position the scans, and a Photoshop action to add an alpha mask and apply radial lightening.
16bit scans worked better than 8 bit scans. Best results were obtained when max,min pixel values were ~90%,10% respectively (much less noise)

The 9 prescanned images I received were also scanned on an Epson and I ran them through the equivalent workflow even though they were only 8bit and exhibited obvious curling on some images.  As they supplied geotiff versions of these as well I was able to redo the georeferencing of the raw scans.  The results I got from Photoscan were better than the supplied georeferenced images.

The biggest problem with the Epson 10000XL (or V700) for large negatives is that you need to find something to flatten the negatives with. Curling negatives can create significant distortion.