FULL-WAVEFORM LIDAR SIGNAL FILTERING BASED ON EMPIRICAL MODE
DECOMPOSITION METHOD
Duan Li, Lijun Xu*, Xiaolu Li, Lian Ma
School of Instrumentation Science and Opto-Electronics Engineering, Beihang University, Xueyuan
Road 37, Haidian District, Beijing 100191, China
ABSTRACT
As a new case of Light Detection and Ranging (LiDAR),
full-waveform LiDAR records the complete waveform of
backscattered echo of targets in certain time interval using
high-speed data acquisition device. Since the full-waveform
signal is generally short in length and badly contaminated by
noise, it is rather difficult to find a method suitable for the
signal filtering. In this paper, the Empirical Mode
Decomposition (EMD) was extended to the filtering of full-
waveform LiDAR signal. Aiming at simulation signal, the
filtering results of EMD-based filtering method were
respectively compared with those deduced from Low-pass
filter, Wiener filter and Gaussian smoothing. The filtering
results show that the Signal to Noise Improvement Ratio
(SNIR) of EMD-based filtering method is biggest in all
compared filtering methods. Residual Sum of Squares (RSS)
of EMD-based filtering method is just bigger than Wiener
filter. Meanwhile, the processing results of different filtering
methods were fitting with Gaussian function using
Levenberg-Marquardt (LM) method. Based on the compare
of fitting parameters accuracy of signal filtered by different
filtering methods, EMD-based method is more suitable for
the preprocessing of Gaussian fitting. At the last, some
typical Geoscience Laser Altimeter System (GLAS) data
were filtered and fitted using EMD-based filtering method
and Levenberg-Marquardt fitting method. The experimental
results suggest that the EMD-based filtering method has well
filtering result.
Index Terms—Full-waveform LiDAR, Empirical Mode
Decomposition, Signal filtering, Gaussian fitting
1. INTRODUCTION
LiDAR is an active remote sensing technique that a laser
scanner emits short infrared or visible pulses towards the
objects and an optoelectronic conversion device records the
backscattered echo [1]. LiDAR measures the round-trip time
of the laser pulse that allows calculating the range between
laser scanner and objects. In combination with accurate
navigation and positioning system, LiDAR can generate
three-dimensional point clouds of object surface. LiDAR
technology has wide applications in geometrics, archaeology,
geography, geology, geomorphology, seismology, forestry
and atmospheric physics [2]. As the development of LiDAR
technique, a new case of LiDAR emerged, which recording
the complete waveform of backscattered echo in certain time
interval by high-speed data acquisition device and called
full-waveform LiDAR. The acquired echo is called full-
waveform signal. Since full-waveform signal contains more
information about measured objects, target features can be
further extracted except the range, such as slope, surface
reflectivity and height.
The most widely used preprocessing method for the feature
extraction of full-waveform signal is Gaussian fitting [3, 4].
As for the Gaussian fitting, filtering is an indispensable
preprocessing step to avoid the noise-induced Gaussian
components and help increase the fitting accuracy. However,
since the full-waveform is generally short in length and
badly contaminated by noise, it is rather difficult to find a
filtering method suitable for the Gaussian fitting. The low-
pass filter, Wiener filter and Gaussian smoothing are the
widely used filtering method, but they all have drawbacks.
Low-pass filter cannot remove the noise with the frequency
lower than the cut-of frequency of the designed filter.
Wiener filter requires to pre-estimate a cross-correlation
vector between input signal and expected signal, which is
rather difficult to carry out because the desired signal is
unknown before filtering [5]. As for the Gaussian smoothing,
the selection of an appropriate kernel width is a key but
difficult work. A small width kernel cannot fulfill filtering
requirement, on the contrary, a wide width kernel will
smooth the features of signal even may deteriorate the shape
of signal. In order to overcome the shortcoming of the
existing filtering method, a filtering method based on the
EMD is presented. This method need not the priori
information about input signal, can improve the Signal to
Noise Ratio (SNR) and reduce Residual Sum of Squares
(RSS) in a large extent compared with other filtering
methods. EMD-based filtering method has strong robustness
aiming at full-waveform signal of different Signal to Noise
Ratio (SNR). More importantly, the Gaussian fitting
parameters of the full-waveform signal filtered by this
method are more accurate.
This research work is supported by the National Natural Science
Foundation of China under Grant No.61201316, No.61121003 and
Specialized Research Fund for the Doctoral Program of Higher
Education (No. 20121102120040)
*Corresponding author: lijunxu@buaa.edu.cn
3399978-1-4799-1114-1/13/$31.00 ©2013 IEEE IGARSS 2013