Summary of First Results

I decided that completing the initial analysis required some sort of closure.  In summary, the StereoDEMs that the PGC are creating are very good.  They are the best thing available right now and that is great.  That said, there is significant room for improvement.  Good thing that both the SETSM, ASP and PGC teams are actively working to refine their techniques.  In particular, some first issues are:

• The coverage is patchy for many of the datasets.
• All methods produce some kind of N-S striping in the difference images.
• Some methods create very rough surfaces.  This is especially evident in the SETSM images
• The differences between the LiDAR and the StereoDEM is not systematic.  Though I have done the first cut analysis to look at the vertical offset, there is obviously some horizontal offset as well.  This will be an avenue of continued work.  Though I had hoped that each type of dataset, regardless of location, would group around a particular vertical offset, this is not the case.  The fact that the distributions shown below have both different mean values and different standard deviations suggests that the lateral offset is potentially complicating things. Note that because the different datasets have different extents and resolutions (hence numbers of pixels), the heights of the peaks are variable and not any an indicator of quality or consistency.
• What is important to note is that any given dataset has a very low standard deviation for the difference between the LiDAR and the StereoDEM.  This suggests that at a particular location, if you knew the true elevation you could apply a vertical offset and be roughly within a few meters of the true elevation. Not bad!

[meters]

First Results: D3 (coalesced/drained lakes)

This is the final post for the preliminary analysis of the DEM data.  This post focuses on D3, the region in the far north of the study area.  There are no moraines here and the substrate is dominated by sands and gravels from glacial outwash.  The lakes that you see are high above the river, back from a bluff edge.

-----------------------------------------Original LiDAR Data-------------------------------------------

Slope-Shaded LiDAR DEM(clipped, projected, ft->m, resampled to 4m)

--------------------------------------ASP_Bay Data-(2 versions)---------------------------------------

Slope-Shaded ASP_Bay_4m_WV02_DEM_11Apr

Vertical difference map (Lidar-ASP_Bay), liner scale between -6m (red) and -9m (blue)

Histogram of above difference map.  Range from -6m (red) and -9m (blue), 

Std. Deviation = +/-2.5m with mean of -7.4m

-----------------------------------------------

Slope-Shaded ASP_Bay_4m_WV02_DEM_12Oct

Vertical difference map (Lidar-ASP_Bay), liner scale between -0.75m (red) and -2.25m (blue)

Histogram of above difference map.  Range from -0.75m (red) and -2.25m (blue), 

Std. Deviation = +/-0.32m with mean of -1.6m

-----------------------------------------ASP_Par Data-------------------------------------------

Slope-Shaded ASP_par_4m_WV01_DEM

Vertical difference map (Lidar-ASP_Bay), liner scale between -0.5m (red) and -2.5m (blue)

Histogram of above difference map.  Range from -0.5m (red) and -2.5m (blue), 

Std. Deviation = +/-0.29m with mean of -1.5m

----------------------------------------SETSM Data------------------------------------------

Slope-Shaded SETSM_DEM

Vertical difference map (Lidar-ASP_Bay), liner scale between 0m (blue) and 3m (red)

Histogram of above difference map.  Range from 0m (blue) and 3m (red), 

Std. Deviation = +/-1.3m with mean of 1.5m

First Results: D2 (region with moraine, streams)

After some focused effort to move manuscripts out the door and present a keynote at a Chilean conference, I am now able to complete the preliminary analysis for the remaining two test areas.  This blog post focuses on D2, the central domain.  Remember there are three test domains were selected within the LiDAR extent (thick red outline) because of their diversity in topographic texture.  D2 is hilly with a well defined stream network.

-----------------------------------------Original LiDAR Data-------------------------------------------

Slope-Shaded LiDAR DEM(clipped, projected, ft->m, resampled to 4m)

--------------------------------------ASP_Bay Data-(3 versions!)---------------------------------------

Three different versions of the ASP_bay DEM data were available in the D2 domain.  I created clipped DEMs and difference maps for each of these.  It will allow comparison of two different time periods (Feb and Oct, snow differences.) and different sensors (world view 1 and 2).  Note, the reason why some of the images do not have full overlap with the LiDAR data is because their extent is limited.  I tried to select a location that provided at least partial coverage for as many datasets as possible.  The fact that the overlap is not perfect is accounted for in the histogram.  Edge effects (relatively minor) are not.

Slope-Shaded ASP_Bay_4m_WV01_DEM_12Feb

Vertical difference map (Lidar-ASP_Bay), liner scale between -5.8m (red) and -8m (blue)

Histogram of above difference map.  Range from -5.8m (red) and -8m (blue), 

Std. Deviation = +/-1.7 with mean of -6.9m

--------------------------

Slope-Shaded ASP_Bay_4m_WV01_DEM_12Oct

Vertical difference map (Lidar-ASP_Bay), liner scale between -5m (red) and -7m (blue).

Histogram of above difference map.  Range from -5m (red) and -7m (blue), 

Std. Deviation = +/-0.75 with mean of -6.1m

--------------------------

Slope-Shaded ASP_Bay_4m_WV02_DEM_12Oct

Vertical difference map (Lidar-ASP_Bay), liner scale between -9m (red) and -12m (blue)

Histogram of above difference map.  Range from -9m (red) and -12m (blue), 

Std. Deviation = +/-4.5 with mean of -10.8m

-----------------------------------------ASP_Par Data-------------------------------------------

Slope-Shaded ASP_Par_4m_WV01_DEM

Vertical difference map (Lidar-ASP_Par), liner scale between -3.5m (red) and -5m (blue)

Histogram of above difference map.  Range from -3.5m (red) and -5m (blue), 

Std. Deviation = +/-0.33 with mean of -4.4m

----------------------------------------SETSM Data------------------------------------------

Slope-Shaded SETSM DEM

Vertical difference map (Lidar-SETSM), liner scale between 0 (red) and -3m (blue)

Histogram of above difference map.  Range from 0m (red) and -3m (blue), 

Std. Deviation = +/-0.96 with mean of -1.4m

First Results: D1 (high relief area)

Now that the automated data processing workflow is very close to complete, I thought I would share some preliminary results from this afternoon's analysis.  These images represent the types of outputs I will generate for the different data sets in the different domains.

Slope-Shaded LiDAR DEM(clipped, projected, ft->m, resampled to 4m)

-----------------------------------------ASP_Bay Data-------------------------------------------

Slope-Shaded ASP_Bay_4m_WV01_DEM 

Vertical difference map (Lidar-ASP_Bay), liner scale between -2m (red) and -6m (blue)

It looks like a northward horizontal shift in the ASP image would improve things.

Histogram of above difference map.  Range from -2 to -6., 

Std. Deviation = +/-1.5 with mean of -4.3m

As Paul says, "Wow."

-----------------------------------------ASP_Par Data-------------------------------------------

Slope-Shaded ASP_Par_4m_WV01_DEM 

(edge due to limited DEM extent)

Vertical difference map (Lidar-ASP_Par), liner scale between -4m (red) and -10m (blue)

Note that there is an East-West horizontal shift issue with this image.  

If the ASP-Par image was shifted west, you would get much better results

Histogram of above difference map.  Range from -4 to -10., 

Std. Deviation = +/-9.1 with mean of -7.8m

These statistics would be better if the image was shifted west.

-----------------------------------------SETSM Data-------------------------------------------

PLEASE NOTE: SETSM DEMs are beta versions not final. 

Upcoming SETSM DEMs will reduce artifacts like the abrupt height change at mosaic boundaries.

Slope-Shaded LiDAR DEM(clipped, projected, ft->m, resampled to 2m)

Slope-Shaded SETSM_2m_DEM 

(note abrupt change in surface roughness at east-west running seam)

Vertical difference map (Lidar-SETSM), liner scale between -0m (red) and -4m (blue)

Note the East-West seam between input images.

South of the seam, there is a clear need for a southward shift in SETSM relative to the LiDAR.  

Hard to say what kind of shift is needed in the upper image.  Same?

Histogram of above difference map.  Range from 0 to -4.

Std. Deviation = +/-1.0 with mean of -1.9m

These statistics would be better if the image was shifted south.

Again,

As Paul says, "Wow."

Making Projections Correct and Consistent

In order to make proper comparison between the different DEMs discussed above, I need to make sure that each of the files being compared has the same projection and the same resolution.  To do this requires a few steps.

Projections:
I have essentially three different data sets in different projections.

• ASP (parabola and bayesaaw) data are in "Albers_Conic_Equal_Area" with an undefined datum
• SETSM data are in Polar_Stereographic
• Airborne LiDAR are in "NAD_1983_Transverse_Mercator"

My first responsibility is to make sure that all the projection information is defined correctly for the files.  With the help of Clarie from PGC I was able to get a definitive answer on which projection each of the files had.  I then needed to define the projections for all the files.  Note that there are 380 tif files that require this treatment.  

I started the process manually, using ArcGIS's ArcToolbox and but found the right-click, batch Define Projection tool tedious as I needed to paste the projection information into each cell of the batch array.  Then I found a better way to batch define projections (or batch anything).  In order to use a recursive search tool (allowing the contents of subfolders to be searched and modified), I put together a simple ArcToolbox 'model builder' routine which allowed me to combine functions to do the job efficiently (thanks Frank Vignati for the idea).  After 2 minutes of building the model, I was able to process all the files within 3 minutes (and have a handy tool for the future).  Someday I'll learn Python and this will be even easier.

I am now confident in all three projections.  They are all now defined as:

• ASP
• NAD_1983_Alaska_Albers, WKID: 3338 Authority: EPSG
• SETSM
• WGS_1984_NSIDC_Sea_Ice_Polar_Stereographic_North, WKID: 3413 Authority: EPSG
• LiDAR
• NAD_1983_Transverse_Mercator, Authority: Custom (defined by mapping contractor)

Consistent Projections?

The next step is to decide if it is better to reproject all the files to have the same projection (over 100Gb of data = slow/tedious) or whether there is a better way.  I think for this first analysis I will use model builder to clip my lidar data to the extents of the 3 domains discussed above, reproject them to match the Stereo DEM data and then resample the LiDAR data to match the Stereo DEM.  Next I will use the extent of the clipped/projected/resampled LiDAR data to clip the Stereo DEM data.  At this point I will have two rasters ready for comparison.  The Model Builder workflow looks a little like this:

Now that a plan is in place...let's see how this all works.

After multiple iterations, I have improved on the model to produce the outputs I am looking for.  I am sure it could be improved (automate adding to map, grouping, symbology), but this is beyond my skills and time.

Clipping Regions for Comparison

For a first comparison, I decided to extract four regions that represent different topographic domains.  In the image below, the three light blue polygons show the domains for D1 (furtherst south), D2 and D3 (furthest north).

Below are descriptions and LiDAR slope-shade imagesfor the first three domains.

• D1: a mountainous, rocky surface with high slopes

• D2: a tundra-covered hilly area with good fluvial network and diffusive hillslopes

• D3: a low relief area with many coalessed thaw lakes, north of the moraines

• D4: a small region with known thermokarst activity between 2009 and 2012
• (this area is not yet defined...considering Horn Lake or Itkillik Riverside features)

??

For each domain I will:

• Extract a subset of elevation data from the tiles with the most coverage.  If there are overlapping tiles from the same data source, I will mosaic those two together (best method? mean? max? largest area?).
• D1
• Aerial LiDAR Data:  
• (will need to be merged before analysis)
• m4_5_6
• m7_8_9_11
• ASP_Par_4m: 
• WV01_12JUN212145259-P1BS_R2C1-102001001B04AA00_WV0
• ASP_Bay_4m_W1:
• WV01_12JUN202207366-P1BS-102001001C17A600_WV01_12J
• WV01_12JUN202207377-P1BS-102001001C17A600_WV01_12J
• ASP_Bay_4m_WV2:
• (No Data in Domain)
• SETSM:
• SETSM_ArcticDEM_46_18_5_3_DEM
• D2
• Aerial LiDAR Data: 
• m18192922
• ASP_Par_4m: 
• WV01_12JUN152224058-P1BS_R4C1-102001001C555C00_WV0
• ASP_Bay_4m_W1 
• (two tiles, same half-coverage, different dates):
• WV01_12FEB132219064-P1BS-1020010017AD8900_WV01_12F
• WV01_12OCT212212221-P1BS-102001001EB61400_WV01_12O
• ASP_Bay_4m_WV2:
• WV02_12OCT022230092-P1BS-103001001B4F9F00_WV02_12O
• SETSM: 
• (strange merge of 2 datasets, one rough, one smooth?)
• SETSM_ArcticDEM_47_18_1_5_DEM
• D3
• Aerial LiDAR Data: 
• m2728293031
• ASP_Par_4m: 
• WV01_12JUN032207416-P1BS_R3C1-102001001CE78000_WV0
• ASP_Bay_4m_WV2:
• WV02_12OCT202307414-P1BS-103001001C901700_WV02_12O
• WV02_11APR062258334-P1BS-103001000A751700_WV02_11A
• SETSM:
• SETSM_ArcticDEM_47_19_1_2_DEM
• D4
• ????
• After extracting the datasets within each domain, I will run a standard set of calculations, assuming that the airborne lidar is the 'most accurate' dataset.
• Straight difference (LiDAR - stereoDEM)
• output map of differences
• look for spatial coherence in differences, suggesting georeferencing errors
• output histogram of differences
• good for comarison against other stereoDEM datasets
• output bulk statistics of differences
• good for comparison between domains

Welcome Post

Hello All.

This Blog will catalog my progress toward evaluating the high resolution stereo DEM generated by the PGC  team (Paul M.).  I will use this blog to share progress toward a full evaluation (horiz and vert; ASP and SETSM) against my North Slope LiDAR data.

The PGC data were recently made publicly available on the Univ Minn website (at: http://www.pgc.umn.edu/elevation/) so it would be best if I did this sooner than later.

Paul and Spencer were able to deliver to me three different datasets for evaluation:

In order to get my head wrapped around all the different files that were sent, I generated some Mosaic Datasets in ArcGIS to organize all the small tiles into groups.  In the end I created these divisions: 

Here are the footprints and densities of the different datasets compared to the LiDAR data (red):

 SETSM DATA: note that data are only available where the shaded relief images are

ASP BAYSIAN DATA: orange is WV2, green in WV1.  Note the underlying shaded relief is not the extent of these datasets, but is just used for reference.

ASP PARABOLIC DATA: Note the underlying shaded relief is not the extent of these datasets, but is just used for reference.

My plan is to select test areas that have different topographic characteristics and compare them against the different datasets.  I have considered testing the following environments/metrics:

• High relief, high roughness (mountainous with bedrock outcrops)
• Low relief, high roughness (lake zone at northern end of LiDAR)
• Watershed delineation, stream profile (middle moraine area)
• Are there areas of thermokarst change that are resolvable between 2009 and 2012? (NE-14?)