The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part Bl. Beijing 2008 
237 
Figure 3. Signal start error histogram. 
The non-tracking method gave a mean error of -22.99m 
compared to the tracking method’s error of -23.41m. However, 
the tracking method has a modal error of 0.91m compared to the 
non-tracking method modal error of 2.23m. The tracking 
method error distribution looks less biased than the non 
tracking; however the long tail of premature triggerings have 
distorted the mean. Ground height is unlikely to vary more than 
30m over a 30m footprint in forests and the maximum height of 
trees can be estimated, therefore a gradual increase of the noise 
threshold to ensure the signal start is not more than the footprint 
size plus maximum tree height with a tolerance from the last 
return may eliminate the long tail. 
These methods perform reasonably well when the ground return 
contains significant energy and is distinguishable form the 
canopy return. In very dense canopies (>90% coverage) little 
signal reaches the ground. In such cases a proportion of the 
measurements will fail; this proportion needs to be quantified. 
This canopy cover is not uncommon for evergreen broadleaf 
forests (Hofton et al, 2002). Accuracy will be sensitive to the 
ratio of the ground signal to noise. 
Figure 4 shows true tree height against derived height for flat 
ground with a range of densities and ages with 5% noise added. 
The five different age classes are visible up the y axis. The 
actual tree height was calculated from the material information, 
taking the range difference between the first leaf return and the 
mean position of the soil returns. This may be slightly different 
to the actual tree height (also recorded) because of the sampling 
of the ray tracer but is the best estimate that can possibly be 
derived from the signal. 
Topography blurs the ground and canopy returns together 
(Harding and Carabajal, 2005), even without noise. Figure 5 
shows a return from an old growth forest on a 30° slope. The 
ground and canopy returns cannot be reliably separated by 
Gaussian decomposition. Undertory will have the same effect. 
Figure 5. Topographically blurred waveform. 
Multi-spectral lidar 
Two wavelengths with different canopy to ground reflectances 
ratios should allow the ground to be distinguished. Figure 6 
shows the ratio of leaf to soil reflectances against wavelength 
for the spectra used. A canopy also includes bark and this, being 
a similar colour to soil will bring the ratios closer to unity. 
There is still a contrast between the visible and near infra-red. 
Wavelengths of 650nm and 800nm have been used for the 
initial trials as they show a large contrast, although an 
instrument using 1064nm and 532nm will be easier to build. 
Hyper-spectral lidar may also be available one day (Kaasalainen 
et al, 2007). 
Figure 6. Leaf and soil spectra and their ratio. 
In the absence of noise the ground occurs at the point of the 
maximum ratio of 550nm over 850nm. Noise complicates the 
issue, unsurprisingly. 
Figure 7. Spectral ratio showing edges due to ground start, with 
noise alongside material contributions. 
If the signals are blurred by topography or understory the 
ground should be visible as a change in the ratio of visible to 
near infra-red reflectance. This can be found with traditional 
edge detection methods.