1 / 35

Return classification

October 2009, Geological Society of America Annual Meeting, Portland, Oregon. Return classification. Ralph Haugerud U.S. Geological Survey c/o Earth & Space Sciences University of Washington Seattle, WA 98195 rhaugerud@usgs.gov / haugerud@u.washington.edu.

bree-brewer
Download Presentation

Return classification

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. October 2009, Geological Society of America Annual Meeting, Portland, Oregon Return classification Ralph Haugerud U.S. Geological Survey c/o Earth & Space Sciences University of Washington Seattle, WA 98195 rhaugerud@usgs.gov / haugerud@u.washington.edu

  2. A lidar point cloud—pure XYZ position profile view 10-ft thick slice 100 ft No vertical exaggeration O 1st return X 2nd return

  3.  1 km ground pointsidentified by semi-automatic processing all surveyed points Nookachamps Creek, east of Mount Vernon, Washington

  4. What is ground? • Ground is smooth • despiking, iterative linear interpretation algorithms • Ground is continuous (single-valued) • No-multiples algorithm • Ground is lowest surface in vicinity • Block-minimum algorithms

  5. Ground is smooth • despike algorithm flag all points as ground repeat build TIN (triangulated irregular network) of ground points identify points that define strong positive curvatures flag identified points as not-ground until no or few points are flagged

  6. Start with mixed ground and canopy returns (e.g. last-return data), build TIN

  7. Flag points that define spikes (strong convexities)

  8. Rebuild TIN

  9. Flag points that define spikes (strong convexities)

  10. Rebuild TIN

  11. Flag points that define spikes (strong convexities)

  12. Rebuild TIN

  13. Despike algorithm • It works • It’s automatic • Cheap(!) • All assumptions explicit • It can preserve breaklines • It appears to retain more ground points than other algorithms

  14. Despike algorithm • Removes some corners • Sensitive to negative blunders • Computationally intensive • Makes rough surfaces • Real? Measurement error? Misclassified vegetation? Cross-section of highway cut

  15. Ground is continuous (i.e., single-valued) • No-multiples algorithm Multiple returns from pulse Single return from pulse

  16. No-multiples algorithms • Fast • Identify open areas • Hopeless in woods

  17. Ground is lowest surface in vicinityblock minimum algorithms • Computationally rapid with raster processing • Tweedy texture • Biased low on slopes • Appropriate block size is inversely proportional to penetration rate • Requires human intervention to adjust block size • Implicit assumption that ground is horizontal (Successful users of block-minimum algorithms work in flat places)

  18. In the real world… • Almost all return classification is done with proprietary codes • Successful classification uses a mix of • Sophisticated code • Skilled human • To adjust code parameters • To identify and remedy problems • Let somebody else do it! and then carefully check their work • We have no useful metrics for accuracy of return classification

  19. Storing the point cloud

  20. The solution • LAS format • Sponsored by surveying industry, esp. ASPRS (American Society for Photography and Remote Sensing) The problem • Data are voluminous and mostly numeric • Binary formats rule! • A standard file format leads to better tools

  21. LAS 1.0 (May 2003) • Public header block • Data set identifiers • Flight day, year • # records • Data offsets and scale factors • Variable length records • Stuff (projection info, …) • Point records

  22. LAS 1.0 (cont.) • Point data format 0 • Point data format 1 • Adds GPS time as DOUBLE (8 byte floating point number)

  23. LAS 1.1 (March 2005) • Header • modified to better identify data that are not direct-from-sensor • Point data • Classification field becomes mandatory • Standard classification values

  24. LAS 1.2 (September 2008) • Complete time stamp on each point record • GPS second + GPS week OR • POSIX time • Per-point image data (RGB), via new point record types

  25. LAS 1.3 (July 2009) • New point data record types to store waveform data • Modifications to header to store pointer to start of waveform data • Flag for files of synthetically-generated data

  26. Tools for LAS files • Fusion • ArcGIS as of 9.3, LAS 1.0, 1.1 … • liblas (http://liblas.org) LAS 1.0, 1.1, 1.2 • Command-line utilities • C/C++ code library • APIs for Python, .Net/Mono • pylas.py(http://code.google.com/p/pylas/) LAS 1.0, 1.1

  27. Anatomy of a lidar data set

  28. What should a data set include? • Report of Survey • All-return point files • Ground points only • Bare-earth raster • First return (highest-hit) raster • Images • Contours • FGDC metadata italics indicate optional elements

  29. Report of Survey • .pdf or .doc or .odt file—or paper! • Data provider, area surveyed, when surveyed, instrument used, processing software and methods, … • Spatial reference framework • Data provider’s report on data quality • Naming, formats, spatial organization of data files Looks a lot like metadata (it is), but in an older and more human-friendly format. The Report of Survey and FGDC metadata commonly have significantly different content. This is a problem.

  30. All-return point files • LAS binary files • Complete time stamp (LAS 1.2+) much better • Organized by tile or by swath Ground points only • Easily(!) extracted from all-return point files, so why bother? • A convenience for AutoCAD community

  31. Bare-earth raster • Format • Many possibilities, ESRI grid is preferable (discuss) • XY resolution (cell size) • Should be a function of return density: ± 1 ground return per cell • Typically in range 2 ft – 5 m • Z resolution • FLOATING POINT! • integer Z requires half the file size, but is almost useless • What about TINs?

  32. First return (highest-hit) raster • Derived shaded-relief image looks like an orthophoto, but with more contrast • 1st-return – bare-earth = buildings, forest • Two ways to construct: • Sample interpolated (TIN?) surface of 1st returns • Bin 1st returns and take highest value in each cell; some cells have NODATA • Better tree and building heights • Can easily see NODATA areas to assess survey completeness

  33. Image files (optional) • Hillshade • Make your own! • Intensity (from 1st returns or ground returns) • A monochromatic low-resolution orthophoto, captured with an active sensor (not dependent on ambient illumination) • RGB orthophotos • A bad idea: drives up cost of lidar by limiting acquisition to mid-day hours

  34. Contours (optional) • You can make your own • See ArcGIS script CartoContours.py • A significant amount of work • Why do you want contours? • Most all analysis is easier with raster (grid) or TIN

  35. FGDC metadata • See recommendations in A proposed specification for lidar surveys in the Pacific Northwest (PSLC website, also included in course materials)

More Related