Kite aerial photogrammetry: 2 the DSM

1st effort. The surface generated from the Sept 2014 block produced a very irregular model, I had hoped contouring would improve the interpretation of the earthwork but it did not. The contour model was horrible, even after taking out the hedge-lines the 1640’s work is barely visible.

I used down-sampled imagery for this model to feed my impatience but I knew from the outset I’d be lucky to map the surface reliably. The weeds were knee high across almost the entire site. It was more DVM (Digital Vegetation Model) than DSM (Digital Surface Model). The contour model was something of a mess:

Earith contoursWind stagger and oblique cover. The Feb 2015 ‘block’ (I think ‘spatter’ might be a better term) has a different problem, the wind was pretty hefty and this means the cover suffered from obliqueness. Despite using a low angled flight an even distribution was not achieved. The wind was from right to left in this view, it almost blew the camera clean off the site:

Wind staggerThe pixel stretched ‘leaning’ trees on the right of the image reveal the impact of the obliqueness as the projection of the image texture is stretched over the surface. The surface however is close to the ground surface (certainly close enough for 1:200) so I decided to commit the 28 hours processing time the the full resolution images demanded for a high density mesh (they were converted to jpeg from the ARW format for ingestion into Stereoscan so some compression occurred).

06020215 1st Ortho Extract 1The ‘first pass’ ortho is encouraging, based on a low density mesh the textured model looks good. A patient wait for a dense point cloud and even a low polygon mesh from it produces sharp geometry:06022015 Ortho4Comparison between the 2 surfaces is revealing:

Earith surfaces ComparedOn the left is the surface derived from 22 uncompressed images processed from a dense point cloud. On the right is the same operation applied to 96 down-sampled images. An object lesson in the ‘crap in- crap out’ effect: processing more images makes no odds if they are down-sampled ones! The surface mesh from the uncompressed imagery reveals the ditch and bank features well, much better than the earlier, vegetation masked, version. The quality of the imagery is crucial, lighting, vegetation level, even the time elapsed between the first and last captured frame all have a bearing on the outcome in terms of shadow and sharpness.

The camera was set up at ISO 1250 in AV with the stop set at f5.6, this gave shutter speeds of between 1/2000th and 1/3200th. These are numbers that are only feasible with a full frame sensor and good glass. The high ISO and shutter speeds are the great advantages of the full frame format- the shots were sharp, nicely contrasty and even using a CPL filter did not present any problems. A little noise is visible in the shadows but that’s a small price to pay for the fast shutter speeds, truly the Sony RX1 is a wonder indeed.

Decimation & scaling the mesh. At a resolution good enough to detect hard detail (gate posts, the Alan Wiliams MG turret and the spigot mortar base) the mesh was way too big for CAD to handle. 2m+ faces is not really workable on even a high spec PC. This was something of a set-back, I wanted to be able to digitise in CAD from the best reso model possible but the limits were hit by the mesh size. I reduced the mesh polygon count to 500k faces  by ‘quadratic edge collapse’ at a 0.5% reduction value in MeshLab and got a model I could pan, zoom and scale….just on the limit of CAD performance. Regen times were awful but I could apply a scale factor to resize the mesh to accord with distances from TST survey…and the digitising can begin.

Digitising the surface for contouring.  Because I wanted to get a selective ‘edge biased’ contour pattern I selected points on the mesh surface manually. This could be automated but after the first effort the selection of nodes by feature looked a better bet to enhance the record of the soft historic features. I upped the point density around the edges of features and dropped it down to skip hedges and trees. The selection process is always where surveying gets interesting and the inherent bias towards the brief of ‘historic record’ or ‘topographic survey’ becomes apparent: selection is based on the end use of the data.

A touch screen interface is excellent for the digitising, with the mesh on a locked layer and snap set to end point it’s simply a matter of tapping away until the density looks ‘about right’.

Contouring. Once a decent size patch was covered the contour model was tested to check the work was on on track:

1st pass contourSmoothing and interval adjustments were made some dud points (hedge and tree heights) weeded out and the plan takes shape: 2nd pass contourthe contours are generated in CAD by TheoContour with settings adjusted for interval smoothing and line type recontoured and checked against the points.

Next: finding the right point density, digitising the break-lines and developing the contour model…

B

About billboyheritagesurvey

Heritage worker
Gallery | This entry was posted in KAP, Survey Practice and tagged , , , , , , , . Bookmark the permalink.

4 Responses to Kite aerial photogrammetry: 2 the DSM

  1. Are you getting coverage of both the north western and south western outworks in your model Bill? At the end of the north western outwork the ditch forms a V shape in the middle of the v there is a shallow mound which could be part of the site.

    • Thanks for taking a look Hamish!

      Well, for the project the intention was to just do the middle bit but I got the Western covered way by ‘accident’ so I’ll have to fly again to get the Southern covered way. On the Magic map you can see the scheduled area is an ‘L’ shape with the fort in the angle. The Southern end is tricky as it’s very wet, with more wires and the historic surface is damaged. The big V shaped ditch is a modern one, the traces of the Southern covered way are faint now as, since the 60s, water levels are managed for wildfowl conservation purposes rather than land drainage and it’s flooded on a regular basis. I’m not sure if I can hold the whole thing up in Photoscan without serious down-sampling so for now I’m working up the 1;200 plan as record in advance of archaeology.

      B

  2. Tom Benedict says:

    Bill, this is obviously a leading set of questions,but they bear asking:

    At the tail end of “Wind stagger and oblique cover” you compare a mesh made from 22 uncompressed frames and a mesh made from 96 down-sized frames of the same subject. The difference is striking, and illustrates, as you say, the ‘crap-in crap-out’ effect. You didn’t say how aggressively you’d down-sampled those images, but at first glance it points toward the advantages of using larger cameras with larger sensors, hosting larger numbers of pixels.

    Do you have a feel for how mesh density and quality varies with resolution? I’m guessing those two image sets came from the same flight, with the camera at roughly the same altitude. Could you recover the level of detail in the uncompressed DEM if you used a lower resolution camera at lower altitude so the ground scale is the same in each set? If so, how would the number of photos required scale with the resolution of the camera and the time necessary to get the additional coverage from lower altitude?

    I’m guessing you can see where I’m going with this. Larger detectors typically come in heavier packages, which in turn require more lift out of the airframe. The cost per pound of lift is a lot more palatable with a kite than with almost every other type of aerial platform. Is this an indicator that KAP may not just be a viable option, but the option of choice for doing aerial photography for 3D modeling of archaeological sites?

    • The answers to the down-sampling rate/ model parity ratio question are a neat little research paper (based on this and the Warham Camp projects) I’m writing. I’ll expand in a blog post when I have got a better grip on the subject. The shape of the answer is, as you suggest, fewer higher res images are better than a bucket load of lower res ones regardless of flying height.

      Photogrammetric processing is (often) optimised for the nadir case so I suspect the results for a low res camera would improve with a regular 60% overlap swath which you can do with a drone but is proving tough with the kite.

      I exported the 1st block as jpeg at 200px on the longest side and from DNG of 6000×4000 which is a drop from 24Mp to 2.6Mp to give them a fighting chance I did them as Tiffs, cutting the processing time in Steroscan was goal and it crunched them in less tha 5 hours, the un-compressed stuff was days in the works.

      I hit snags further down the process path with the CAD polygon limit so the mesh was decimated just to get the advantage of CAD handling.
      Luckily my ‘full frame’ camera is only 480g so flying it off a kite line is no sweat: but its still a big pay load for a hobbyist class drone.

      When KAP works it’s a really big win, when it doesn’t and the drones get the cover that’s the way of the world. What I want to show is that KAP shouldn’t be over looked as an option before the trigger finger goes for the drone, you never know, you could cut project costs massively and get better res results!

Leave a Reply

Fill in your details below or click an icon to log in:

WordPress.com Logo

You are commenting using your WordPress.com account. Log Out / Change )

Twitter picture

You are commenting using your Twitter account. Log Out / Change )

Facebook photo

You are commenting using your Facebook account. Log Out / Change )

Google+ photo

You are commenting using your Google+ account. Log Out / Change )

Connecting to %s