9-71 Nov. 1999
(cluded from processing
istics.
es for peak positions,
based on the first and
signal, with exception
s.
on fitting to obtain first
tudes, with exception
1s. Editing peaks based
)ximity (zero amplitude
exception handling for
icance if necessary, and
ocation, and half-width.
> start of the waveform
peaks, 2) centroid of the
fined as follows:
(2)
(3)
ting the model, and the
ude
Y; is the observations
the 1-sigma uncertainties
; are calculated in the
(4)
in Y, then the total chi-
(5)
fit Gaussian distributions
dded, user-supplied IDL
ardt, NASA/GSFC Code
Aarquardt technique as a
h for the best fit in the
International Archives of Photogrammetry and Remote Sensing, Vol. 32, Part 3W14, La Jolla, CA, 9-11 Nov. 1999
parameter space (Bevington and Robinson, 1992). A Gaussian
distribution is used because SLA does not digitize the shape of
the transmitted pulse. Lacking information on the shape of the
transmitted pulse on a per shot basis, a Gaussian distribution is
used as a reasonable approximation. A non-Gaussian, user-
supplied fitting function can be input to MPFIT. Updated
versions can be found on http://astrog.physics.wisc.edu/-
-craigm/idl.html. ^ MPFIT uses the Levenberg-Marquardt
technique to solve the least-squares problem. Within certain
constraints, these routine will find the set of parameters which
best fits the data in the least-squares sense; that is, the sum of
the weighted squared differences between the model and data is
minimized by minimizing the Chi-square value, calculating
derivatives numerically via a finite difference approximation. A
set of starting parameters are specified, and the user can apply
constraints to individual parameters by setting boundaries on
the lower and/or upper side. Otherwise, it is assumed that all
parameters are free and unconstrained. The step size to be used
in calculating the numerical derivatives of the model with
respect to the parameters can be defined by the user or it is
computed automatically, and the covariance matrix can also be
computed.
The following constraints are applied during the fitting search:
1) lower bounding constraint that all return position, amplitude
and width parameters be non-negative; 2) upper and lower
bounding constraints on each peak position such that its final
position can not be outside the estimated position bounded by
its half-width (stopping peaks from migrating away into larger
adjacent peaks); 3) if detector saturation occurs, only that part
of the signal up to the saturation point is fit, the estimated peak
position is seeded earlier in time to compensate for the falsely
broadened return, and the area under the saturated peak is
preserved in the Gaussian fit; 4) if the digitizer 8-bit dynamic
range is exceeded causing clipping of the signal, then the signal
before and after clipping is fit making up for the clipped part of
the signal with reasonable accuracy; 5) shots with more than 10
peaks were excluded as they were difficult to fit, usually lacking
convergence after 40 iterations (these are probably low-lying
cloud returns mis-classified as ocean or land surface returns).
4.2. Deriving Elevations from the Waveform
The SLA ranging electronics provide the range from the start of
the transmit pulse to the start of the backscatter return. It is this
range that is used in the geolocation processing and, thus, the
location of the geolocated bounce point (latitude, longitude,
and elevation) refers to the highest detected feature within the
100 m diameter laser footprint. The distances from the start of
the waveform signal to the centroid of all the peaks, the
centroid of the last peak, and the end of signal, all provided in
the SLA-02 data structure, can be used to correct the bounce
point elevation depending on the character of the return signal
(e.g., single or multi-peaked), the assumed nature of the surface
type, and the intent of the end user.
For a measure of the mean elevation of illuminated surfaces
within the laser footprint (assuming uniform reflectance of all
the surfaces at the 1064 nm laser wavelength) the appropriate
range would be from the centroid of the transmit pulse to the
centroid of the backscatter return. An approximation of this
mean elevation is obtained by:
[geolocated bounce point elevation] + [transmit pulse
centroid] — [centroid distance for all return peaks] (6)
The transmit pulse centroid corresponds to the distance from
the start to the centroid of the transmit pulse. SLA does not
provide a digitized record of the transmit pulse. However, the
transmit pulse impulse response can be obtained from returns
from flat, smooth surfaces such as water, defining the narrowest
possible returns which have not been broadened in time by
surface relief. The impulse response pulse width is a function
of peak amplitude. Examination of a plot of waveform centroid
versus peak amplitude for single-peak returns defines an
envelope of data points, with the minimum boundary defining
the impulse response centroid which varies from 2 meters for
low amplitude returns to 4 meters for high amplitude returns.
For multiple-peaked waveforms where it is assumed that the
last waveform peak is due to laser energy backscattered from
the ground and that preceding peaks are energy returned from
higher surfaces such as vegetation or buildings, the mean
elevation of the ground surface can be approximated by:
[geolocated bounce point elevation] + [transmit pulse
centroid] — [centroid distance to last peak] (7)
Inference that the last peak corresponds to the ground surface
within a 100 m diameter footprint requires that the height
distribution is simple, as for example due to an open vegetation
canopy or building above a flat ground surface. Ground slope
across the footprint can cause convolution of the ground return
with returns from overlying surfaces.
For a measure of the lowest detected surface within a footprint,
the appropriate range is the distance from the start of the
transmit-pulse to the end of the waveform signal, with a
correction for the full width of the transmit pulse impulse
response. This elevation can be approximated by:
[geolocated bounce point elevation] + 2%*[transmit pulse
centroid]— [distance to end of waveform signal] (8)
These calculations assume the laser vector is at nadir. For off-
nadir pulses, the distances along the laser vector to the
centroids and end of signal should be modified by multiplying
by the cosine of the off-nadir pointing angle, although this
modification is very small for the near-nadir SLA observations.
The waveform processing and resulting products thus provide a
means to correct the SLA first return elevation to a mean
elevation for illuminated surfaces, a mean elevation of the last
return, and the elevation of the lowest return. The appropriate
use of these elevations will depend on the assumed character of
the surface within the laser footprint and the intent of the user.