An Automated, Open Source Pipeline for Mass Production of 2 m/px DEMs from Commercial Stereo Imagery

Video in TIB AV-Portal: An Automated, Open Source Pipeline for Mass Production of 2 m/px DEMs from Commercial Stereo Imagery

Formal Metadata

An Automated, Open Source Pipeline for Mass Production of 2 m/px DEMs from Commercial Stereo Imagery
Title of Series
CC Attribution 3.0 Germany:
You are free to use, adapt and copy, distribute and transmit the work or content in adapted or unchanged form for any legal purpose as long as the work is attributed to the author in the manner specified by the author or licensor.
Release Date
Open Source Geospatial Foundation (OSGeo)
Production Year
Production Place
Portland, Oregon, United States of America

Content Metadata

Subject Area
We have adapted the NASA Ames Stereo Pipeline (ASP) - a suite of automated, open source, command-line photogrammetry tools originally developed for NASA planetary missions - to process high-resolution stereo satellite imagery of the Earth. These tools are multithreaded, memory efficient and scalable, which enables processing of "big image data" (e.g., 16-bit panchromatic WorldView images with dimensions ~36000 x 460000 px). We have deployed this pipeline on the NASA Pleiades supercomputer to generate ~2 m/px digital elevation models (DEMs) and ~0.5 m/px orthoimages for thousands of WorldView-1/2 along-track stereopairs. New ASP tools mitigate systematic DEM artifacts and allow for automated, a posteriori DEM coregistration using iterative closest point algorithms. When existing control data are available (e.g. LiDAR, GPS), automated alignment routines offer sub-meter horizontal and vertical DEM accuracy.Our research applications focus on ice sheet dynamics in Greenland/Antarctica and ice/snow evolution in the Pacific Northwest. We have developed an additional collection of tools for DEM analysis, including utilities to produce maps of 3D surface displacement (velocity) vectors and eulerian/lagrangian elevation change. We present the following case studies to highlight the capabilities of these data and our open source workflow:-A 57+ DEM timeseries from 2008-2013 for Greenland's most dynamic outlet glacier, revealing >40 m/yr interannual thinning and large seasonal variability-Annual DEM mosaics that reveal the ongoing evolution of West Antarctica's "weak underbelly", an area roughly the size of New Mexico-Repeat DEM timeseries for Mt. St. Helen's showing volcanic dome growth, glacier advance, canopy height, fluvial erosion/deposition, and landslides.For many applications, DEMs derived from high-resolution satellite imagery are comparable to those derived from airborne LiDAR data, with the advantage of global, on-demand tasking capabilities and reduced costs. Archived commercial stereo imagery is available at no cost to federal employees or federally-funded researchers, and the tools/methods highlighted here offer an automated, open source alternative to traditional, GUI-based, commercial photogrammetry software packages.
Keywords photogrammetry remote sensing 3D DigitalGlobe WorldView
Single-precision floating-point format Projective plane Mass Image registration Number
Metre Area Satellite Slide rule Uniform resource locator Pixel Image resolution Surface Cuboid Water vapor Analytic continuation Orbit
Metre Dialect Image resolution Surface Motion capture Set (mathematics) Insertion loss Water vapor Student's t-test Measurement Formal language Mathematics Personal digital assistant Task (computing) Spacetime
Satellite Dialect Context awareness Stereographic projection Operator (mathematics) Polygon Green's function File archiver Mereology
Polar coordinate system Point (geometry) Polynomial Open source Multiplication sign Workstation <Musikinstrument> Gene cluster Set (mathematics) Modulare Programmierung Computer Computer programming Machine vision Power (physics) Linker (computing) Core dump Graph (mathematics) Endliche Modelltheorie Rational number Binary code Volume (thermodynamics) Bit Website Quicksort Coefficient Directed graph Library (computing)
Texture mapping Arm Wrapper (data mining) Tesselation Line (geometry) Directory service Parameter (computer programming) Function (mathematics) Product (business) Digital Revolution Term (mathematics) Authorization Point cloud Utility software Series (mathematics) Endliche Modelltheorie Error message
Metre Point (geometry) Dialect Pixel Game controller Texture mapping Multiplication sign Genetic programming Plotter Range (statistics) Workstation <Musikinstrument> Set (mathematics) Function (mathematics) Line (geometry) Number Product (business) 10 (number) Goodness of fit Term (mathematics) Different (Kate Ryan album) Operator (mathematics) Vertex (graph theory)
Metre Rotation Point (geometry) Covering space Addition Shift operator Spline (mathematics) Plotter Sampling (statistics) Translation (relic) Function (mathematics) Image registration Rigid body Demoscene Product (business) Number Fraction (mathematics) Uniform resource locator Network topology Point cloud Utility software Error message
Point (geometry) Area Metre Standard deviation Spherical cap Schmelze <Betrieb> Right angle Line (geometry) Stack (abstract data type) Marginal distribution
Frequency Population density Multiplication sign Time series
Metre Frequency Pixel Uniform resource locator Multiplication sign Mathematical analysis Linearization Insertion loss Musical ensemble Twitter Social class
Multiplication sign Product (business)
Metre Pixel Focus (optics) Causality Insertion loss Mass Mereology
Metre Area Complex (psychology) Mathematics Texture mapping Discrete element method Population density Right angle Product (business) Social class
Trail Uniform resource locator Mathematics Texture mapping Scaling (geometry) Multiplication sign Schmelze <Betrieb> Insertion loss Mass Spacetime Product (business)
Trail Pixel Discrete element method Multiplication sign Surface Image resolution Mathematical analysis Product (business) Derivation (linguistics) Uniform resource locator Vector space Velocity Single-precision floating-point format
Dot product Uniform resource locator Concentric State of matter View (database) Multiplication sign Polygon Water vapor Quicksort Office suite
Multiplication sign Product (business)
Revision control Concurrency (computer science) Cross-correlation Term (mathematics) View (database) Image resolution Network topology Zoom lens Product (business)
Metre Discrete element method Term (mathematics) Different (Kate Ryan album) Right angle Mass Form (programming)
Metre Noise (electronics) Word Mathematics Texture mapping Forest Network topology Multiplication sign Bit Measurement Graph coloring
Mathematics Volume (thermodynamics) Measurement Arithmetic progression Near-ring
Arithmetic mean Mathematics Resultant Data quality
Metre Point (geometry) Game controller Pixel Service (economics) State of matter Multiplication sign Image resolution Time series Temporal logic Computer programming Product (business) Photographic mosaic Population density Automation Endliche Modelltheorie
Metre User interface Area State observer Discrete element method Information State of matter Surface Multiplication sign Image resolution 1 (number) Maxima and minima Function (mathematics) Cartesian coordinate system Mereology Product (business) Term (mathematics) File archiver Utility software
married I I think we should value started here it's 3 o'clock I and so I have a lot to talk about and very excited about a lot of things on the show here and my name is David Sheena mysteriously Washington but working with the number of people who have been instrumental for this project include Claire was from and so this
document gives the background of my talk about the the massive amount of data that we have to talk about our processing tools a single registration accuracy analyses that I've done and then show a lot of pretty pictures of which I think you guys will enjoy verb places all over the planet and so this is my
1 slide summary of what we're doing here so essentially we're taking stereo satellite imagery and primarily were using DigitalGlobe worldview 1 world to data a worry these these satellites are incredibly nimble and precise in a single orbit they can acquire an along-track stereo so they can pitch forward acquired image and then about a minute later pitch back an acquired image of the same spot on these images are half meter per pixel when we receive them as a 17 km swath so it's 17 kilometers by however many kilometers along the other dimension and we've developed a pipeline to take those images and automatically produce high-resolution continuous digital elevation models and they're doing this over and over and over again for some high priority areas but this location agreement so what is half meter per pixel
images actually look like and this is a an image acquired in Greenland and 2011 This is the Greenland ice sheet here was black things in the inset there those Earth's surface melt water lakes and in that little red box is this larger image here you can see there's a stream channel here some of these darker areas there's no water on the surface and is the strange bumps and those are
actually mountain tense and these 8 by 8 foot photons from the 2011 Woods Hole and camps sets kind resolution we're working with here and I'm a PhD
student my I'm in the Earth space sciences department so I'm task doing science so the motivation for a lot of work that I'm doing here is that these repeat measurements of surface topography can capture ice loss and in some cases I skiing but but really they characterize the processes that are driving a lot of changes that were seen today on the ice sheets and elsewhere we care about this because all that ice that's leading the ice sheets ends up in the ocean and every every year sea levels rising 3 mm year about 1 millimeter of that is from the ice sheets we wanna know where that's going on the future we know it's going increased we don't know how much we can also learn things about regional climate change by studying languages and something that is so talk earlier really hammered home and we can if we know elevation of seasonal snowpack we can learn a lot about water resources which is really critical for the western United States and so in terms of
coverage here's what avail is available as of June 1st for it in the difficult archive every polygon here is a along-track stereo pair so there's 175 thousand these roughly and and we look at this you see there's excellent global coverage but really the polar regions are extremely well covered by some and part of the reason for that is that there is now 6 small worldview 3 is an operation 6 polar orbiting satellites every 90 minutes 1 of these is flying over 1 moles and so there's a lot of opportunities this also essentially 0 competition the Virtue customers that want to take a picture in the middle of an article and painted go to do that answer fortunately the federal government has an arrangement between National Science Foundation National Geospatial-Intelligence Agency few other agencies to make these data available to researchers I guess so these the
polls how many people actually use of used a polar stereographic projection just curious OK that's more than I was expecting from OK so this is an article in green and everything in blue is a mostly cloud-free stereo-pair acquired in the last 4 years everything in red is what I have received so far from and just put inverse were talking about 10 thousand of these pairs in Greenland and roughly 20 thousand article so huge amounts of data here and and just for a context
if you go to map Flight . com and you've pick an order covers the United States of America is 1 and a half times the size United States including Alaska and so to put this is we're talking about sigh you know the common US plus so clearly
some you know gooey based techniques to process the data are not gonna work you only have this volume and so what we've done is we've partnered up with some folks at the NASA Ames Research Center the Intelligent Robotics Group there and they have an existing pipeline called the Ames stereo pipeline that's a set of automated open-source command-line tools and he's originally developed for a Mars Pathfinder back in 1987 and and then in about 2008 they adopted to process pretty much any image taken with any cameron assets flown in the solar system sold up here we have data from the Apollo mission and these polymetric camera and this is a mosaic I put together for this is Gale crater on Mars this is the Mars Science Laboratory curiosity landing site here and so in 2012 received funding from the NASA cries your program to adapt these tools to process these Earth data and so at this point we supported is a little rigorous camera model on any camera model that has our PCs user rational polynomial coefficients DigitalGlobe distributes them the Airbus the treaties data they they have the power PCs and skybox is doing and I believe time-lapse has our PCs as well Frank is sort of not it OK and so all of the all these different images can be processed with this these tools here and the core is written in C + + it's mostly based on a custom computer vision library that's built to handle huge images in a very efficient way and most of the core tools a the memory efficient and this these tools are built to run on clusters and so fortunate I have quite a bit of time on the NASA please cluster which is enabled me to really do things that would take months or even years you want just to local workstations or smaller clusters you can download binaries of this software package right now for Linux and OS 10 and the source is always available and it not give you links later
so in terms of how it actually works and it can be very simple there a series of Python wrappers that 1st line of their stereo image 1 image to an output directory that on the entire pipeline and spit out point cloud for you and the 2nd line is essentially a grading utility that takes the point cloud and turns it into an elevation idea with syntax very similar to the G all utilities and or you can get in there and 2 as many parameters as you'd like and so it's highly customizable which is great because you can it's it's good and bad from
as we all know so it in terms of how the pipeline actually works up there on the left and the data are delivered in smaller tiles 1st thing we do is correct the the original images to deal with some slight Slade not errors by the major shortcomings in the digital camera model and then we mosaic them together we doing orthorectification so the utilities here for the orthorectification multicellular orthorectification arm and then we generate a series of disparity maps so to match features between the 2 images that we find them we filter them up at a point cloud and then we output of DM and final author rectified products
but most people want here really care about these things and so we're talking about half an hour 24 hours on a single dedicated workstation so and that range is due to the fact that there's a lot of different settings but if you want a low quality product it'll take you half an hour if you want the highest quality product possible you looking at somewhere in the range of 4 to 12 hours maybe even 24 hours also depends on the actual image data if here you have a lot of texture and features in image it's can run much faster than it would otherwise and in terms the output products were posting news about 2 meters per pixel and in terms of accuracy are uncorrected vertical and horizontal accuracy this is just added the canned from DigitalGlobe about 5 meters vertical and 5 meters horizontal however is that a lot of time in the past few years bringing the number down and I can say with confidence that if we have good control data we can get below to sub meter horizontal and vertical fortunately for the polar regions we do
have a lot of good control data and so NASA every summer is flying multiple aircraft to each bowl I had this is really cool mission and it's called Operation IceBridge and have multiple light are instruments that are providing a lot of data like those orange line Singapore plots with 10 centimeter accuracy was also existing data from the laser altimeters that I sat mission and we have GPS campaigns so we have a lot of points that I trust within tens of meters
but 1 of the trees in additions to the hamster a pipeline utility to do automated point-cloud point cloud cover registration so basically this is an iterative closest point alignment where you have 1 point cloud here if you the 1 that's close but not exactly aligned and you can do a rigid body translation and rotation to fit the to minimize the errors and so on the left there we have is a the degraded DM product that's and corrected by the middle panel of clipped outlined our data over bedrock for this particular location and then we run this utility and the output medium there is roughly 45 centimeters for all those locations after 3 and a half meters shift
so there's a lot here and um I keep moving on but basically the the scene idea nearly 90 numbers over there but those are the numbers that I'm getting for 31 diems over the same piece of bedrock agreement so that's roughly what DigitalGlobe spec it so that's good but I believe that and that this plot here on this basically showing so each blue and red pair is the before and after the sampling of all of my control points for each and so you can see that before I do that fraction I have errors of you know 4 5 meters and then afterwards everything is down here about 7 years and if you just a
believer this point even more so basically this is that same stacks of got 31 DMEs covering this 1 piece of rock um and on the right here is the standard deviation of elevation values and so on some of these red areas are places where you have snow accumulating melting and there's lakes and things but these areas in here that's just bedrock and we can see that we have roughly and of 30 to 40 centimeters standard deviation so we really are a sub meter what this does is if I trust these elevations to 30 to 40 centimeters like can trust elevations over the ice with the same kind of uncertainty OK so Stanford
pretty pictures and also this is Greenland and this is the West's margin of the Greenland ice sheet and this is a mosaic that is 220 by 240 kilometers from and where these 2 dash lines meet that's the terminus of Yaqub shopping is brain it's the fastest glacier on the planet it's moving about 18 kilometers a year which turned that's 2 meters an hour so it's not typically what we think about we think of clashes on its dumping tremendous amounts of of ice in the ocean and actually it is fun fact this is and believed to be the glacier cap the iceberg sank the titanic so produces the huge kilometers sized icebergs and and it's changed a lot in the past
few decades so we have annual coverage here and and really what's
exciting to me is we have incredibly dense time series of for the terminus the very calving front so I have 46 of these worldview diems in this 5 year time period from 2008 to 2013 and
i've absence supplemented with Tandem-X The German X-band SAR mission also beautiful products but basically you've got you 57 in these TMs we can bring them together
and start to do some analysis here there's a lot on this slide but on the left what I'm showing is basically a linear elevation trend every single pixel at this location and so what you see is that the last year you alterations and it the but it is certainly not the end of the day the this issue that that's 1 of the the the starting from the what kind of music as a lot of old age class his 5 year time period of about 100 to 150 meters of loss at that location and on top of that what we see is a very strong seasonal cycle without by 1 of this it's only going up and down 30 to 50 meters season Mr. never really been observed for up so pretty cool but but I wanna move on this is a movie
that I threw together and showing
what happened in the past 5 years this glacier
and and you know it's these are preliminary
products it's way too fast and you know
you guys are gonna take giving away from this but it's
it and put in here just cause it's really cool but
also in were effectively able in near real
time to observe what these glaciers are doing and that's that's impressive we never been able to do this
for a OK ominous which polls on
you know this is an article and how many people saw the press release maybe 2 months ago I'm about unstable collapse of the Western irrigation 9 and parentheses it's happening the cause of focus so if you that's happening right here this is this is the part of Western article that everybody's worried about because it's it's changing very rapidly and it's in a situation where there's really nothing that's going to stop a lot of loss that were seen there right now I'm so fortunately have some excellent coverage here from the past 4 years and this is a mosaic uh put together this is 520 by 690 kilometers so that's roughly the size in Mexico want to put this in scalar mass at 2 meters per pixel and then for reference this is the best available
data this bad map to DEM was produced about 2 years ago and it's good it's 1 of 5 corners pixel but you know when
you compare to something like this there's some significant areas hundreds of meters and
so this is a big change for us and and this these new maps are
basically enabling us to produce products like this and what I'm showing you this is now a relative to sea level so 0 is roughly sea level and all all right is around the size on well this is the thing that where the the the the the the the 1st thing to note is that the chance that the that's the way the way we can identify is because this all this in values and so assuming that class so we're only seeing basically the upper 10 per cent of the whole complex so for measuring this we can infer this because we know the density so what what this allows us to do
is over time again we have a year for every year we can basically track how the ice is thinning but also how the ice shelves the melting of over time and this is just a
preliminary product showing some of the elevation change maps for this location this is the Pine Island Glacier and over here this is the previously real data the user mass I sat satellite altimetry tracks you could see this this scaling of the the problem the spacing between a lot of these tracks is 10 to 30 40 kilometers and so we got an idea that there was the mean going on here and elevation loss but now we can actually pinpoint exactly where that's occurring when it's occurring and how much current so very exciting and ences
stadium and yeah I'm at 1640 um of briefly mention these products are high enough resolution that we can actually use the actual DEM data to derive surface velocities so this is essentially running feature tracking and we're using an image correlated to track every single pixel in 1 DM at time 1 to its new location in time to what that gives this is a complete 3 dimensional vector for every single pixel from time 1 time 2 time 3 time 4 of and this is incredibly valuable and from ecclesiological standpoint so very excited about this and its enabling us to do
all kinds of new analysis and I don't wanna get into the details here because I wanna show you guys this
time uh so I live in Seattle now and I spend a lot of time and these mountains and in the cascades of it's beautiful place to go hiking and and uh are office we have a sort of a view of the top of Mount Rainier and sitting there 2 years ago and same as of why we not collecting these data for the mountains here in the ice in Washington in Washington state has the highest concentration of ice of any state in the lower 48 from Alaska you know these blows Washington water but and it basically are using these these blue polygons is where there's a glacier on each 1 was red dots is where we put in a target to acquire the stereo data and you can see these postage stamps so the locations where we got them in 2012 and 2013 so we we did get some usable data unfortunately from a lot of
our data looks like that and this is the Pacific Northwest and it's cloudy here but but we did get some good data for
the St. Helens and so is some early products put together for announcer 2012 and 2013
and this is the bearer light data for St. Helens is a really nice dataset from 2003 and back before our was really very prevalent in time and this is the
world view stereo products and now I'll go back and forth just you guys
concurrency were doing pretty well and in terms of resolution and also in terms
of alignment accuracy and a lot of these these no data gaps here those are actually trees were the correlation failed and the more recent versions the pipeline actually fill in a lot of those but I didn't have a chance to make his slides and in just a zoom in
to really hammers home the on the left is the light on the right is the worldview stereo DEM and the quality in terms in terms a resolution and in the south side and we were able to resolve things there's several meters across and if you look at these you can see that things are changing their that's in houses when the most dynamic places on the planet right now and and this is a lot of weight here is that form there so we can look at the
elevation differences in there is roughly 200 meters of new Rockford erupted during the 2004 to 2008 eruption there and the glacier which and want the the and call us and blowing up north side of the crater and so this is kind 1 of the few success stories for glasses in are in the world that we live in right now is that this is what I think is really cool I'm so if
I if we step back a little bit saturate those changes we begin to see all kinds of other things so that when I 1st made this map I saw these these horrible colors around the outside and inside of a problem but it would I realize that this 2003 light our data it's bearers data so they've removed all of the canopy all the vegetation and where word with our stereo images were actually correlate on the tops of the trees and so this is actually a measurement of vegetation which for me is noise not interests but for a lot of people that's it that's a signal like they're very interested in vegetation height over time Especially here where you lost most of your forest in 1980 from but you use Conner saturated you can really see it but he that's that this is a lot of people the me of the have the signal there if i is bent over 20 meters of erosion
along this river in the past 10 years and actually lie deposition up in the left corner there you have these meanders and then there's a that's a small is known the servatory I'm so this just kind demonstration of what's possible with these data we were using them to make
quantitative measurements of volume changes of ice that's measure glacial advance and then to measure snowpack on both winter accumulation in winter snow melt her summer so
but I and just a progress update that we've got 15 targets now and the 2014 acquisitions have been going very well we're getting data from April May and then again in September and October as well as monthly stereo near um and then
we recognize that the but it has not had and so what this is everything there so what the mean and the literary but Data Quality amount that is is excellent I don't have any results for changes yet because this is the only data that I have but here were in there
somewhere and and there's not heard and just to give you a sense of the the quality these
products from the just in here for the gratuitous 3 rendering them and there yet so is
summarize we've got this automated pipeline time it's working very well were producing tuneup pixel elevation models with 5 meter uncorrected accuracy and sub meter corrected accuracy with really you know I can do this with just 9 ground control points it doesn't take much from the way I see it these repeat PDM's offer unprecedented spatial and temporal resolution to study dynamic processes and and I see at each of these is kind of like an airborne lighter service not can replace light are but it's it's becoming close and I think we can learn a lot especially over rock and ice with these data and also these when you have these dense time series of I can get the state every week it's awful lot like having GPS receivers every 2 meters and enables you to do a lot of things that you wouldn't be able to do otherwise our generating mosaics and were never be generally mosaics there over a million square kilometers and keeping the baselines for these work are established this point and you were just starting this program and my only I hope that we're going to continue begins at 10 or 20 years from now on so announced that and here's
the info and I'll stop there and I encourage you did to download and check out this package it's it's cool there some useful utilities in there and I'm happy to take any questions now or after the session thanks if
the but the was that the L 1 of the son of there's a target area there we we got some data for us the Olympics have even know making any products serve using metal and so that we this whole on the part of the and what so many the things that he did you know that the the the convention to that process of this on the other hand we do not know the difference of about how they did the Dems get 1 what and processing them in terms of size of size but it depends on the final output resolution scientific and they compressed pretty well you so for a postage stamp like this were probably talking about a few hundred megabytes but if we go back to some of these other ones you know each of these strips that you want so I think these things the state thank yeah that and he thought about using this in forestry applications such as you have like a bear and then stand height or whatever have thought about it but I would have gotten us to work on already I I I have started talking some folks at Goddard who were really interested in those products of high you know this is only beginning I'm sure that these will be used forestry applications the can you go back through the Digital Globe archive and use any pairs so you if they weren't flown as the just the other yeah and so like a coincident stereo-pair pair yet but I can do that assuming that my surfaces and changing so for places like the cascades that's usually I can take an image acquired on 1 day in an image acquired maybe a week later and I can probably generate about the way but for some more like Yokuts of his brave it's moving 2 meters an hour I can't my maximum time window there's pride 6 to 12 hours and of dried but I have actually made some stereo pairs from non along-track observations and they've worked through all so the why did you do in the future


  501 ms - page object


AV-Portal 3.20.2 (36f6df173ce4850b467c9cb7af359cf1cdaed247)