Transcript 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
[email protected] / [email protected]
A lidar point cloud—pure
XYZ position
profile view
10-ft thick slice
100 ft
No vertical
exaggeration
O
X
1st return
2nd return
 1 km 
all surveyed points
ground points identified
by semi-automatic
processing
Nookachamps Creek, east of Mount Vernon, Washington
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
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
Start with mixed ground and canopy
returns (e.g. last-return data), build TIN
Flag points that define spikes
(strong convexities)
Rebuild TIN
Flag points that define spikes
(strong convexities)
Rebuild TIN
Flag points that define spikes
(strong convexities)
Rebuild TIN
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
Despike algorithm
•
•
•
•
Cross-section of
highway cut
Removes some corners
Sensitive to negative blunders
Computationally intensive
Makes rough surfaces
– Real? Measurement error? Misclassified
vegetation?
Ground is continuous (i.e., single-valued)
 No-multiples algorithm
Multiple returns from pulse
Single return from pulse
No-multiples algorithms
• Fast
• Identify open areas
• Hopeless in woods
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)
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
Storing the point cloud
The problem
• Data are voluminous and mostly numeric
Binary formats rule!
• A standard file format leads to better tools
The solution
• LAS format
– Sponsored by surveying industry, esp. ASPRS
(American Society for Photography and Remote
Sensing)
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
LAS 1.0 (cont.)
• Point data format 0
• Point data format 1
– Adds GPS time as DOUBLE (8 byte floating point number)
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
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
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
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
Anatomy of a lidar data set
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
italics indicate
FGDC metadata
optional elements
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.
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
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?
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
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
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
FGDC metadata
• See
recommendations
in A proposed
specification for
lidar surveys in
the Pacific
Northwest (PSLC
website, also
included in
course materials)