Infrared Physics & Technology 130 (2023) 104613

Contents lists available at ScienceDirect

Infrared Physics and Technology

journal homepage: www.elsevier.com/locate/infrared

Full-waveform LiDAR echo decomposition method based on deep learning

and sparrow search algorithm
Xiaobin Xu *, Jiali Wang, Jialin Wu, Qinyang Qu, Yingying Ran, Zhiying Tan, Minzhou Luo
College of Mechanical and Electrical Engineering, Hohai University, Changzhou 213022, China
Jiangsu Key Laboratory of Special Robot Technology, Hohai University, Changzhou 213022, China

A R T I C L E I N F O A B S T R A C T

Keywords: For the full-waveform LiDAR echo, the conventional decomposition method is to judge the number of waveform
Lidar components in the echo through the IP (inflection point) method or RL (Richardson-Lucy) deconvolution method
Full-waveform decomposition after filtering and obtain the initial estimation of their parameters, and then use the intelligent optimization
Sparrow search algorithm
algorithm to fit the waveform. However, this preprocessing method has poor accuracy in judging the echo
LSTM
components when the distance between targets is small. In this paper, a highly accurate decomposition method is
proposed based on LSTM with SSA (Sparrow Search Algorithm). Firstly, the LSTM network trained by the
simulation data sets under three kinds of background noise is used to judge the number of Gaussian components
in the full-waveform LiDAR echo, and then SSA is used for waveform fitting. The accuracy rate of each LSTM
network is more than 95%. This method is compared with IP, RL and MGD (multi-Gaussian decomposition)
method. Within the accuracy of 0.1 m, the minimum decomposition distance of LiDAR echo is shorten from 0.7 m
to 0.45 m compared with IP method and from 0.55 m to 0.45 m compared with RL and MGD method. The
minimum ranging distance of LiDAR echo is shorten from 0.75 m to 0.65 m compared with IP and RL method and
from 0.85 m to 0.65 m compared with MGD method. With proposed method, the accuracy of full-waveform
LiDAR echo decomposition is improved and distance limit of decomposition is reduced.

1. Introduction parameter estimation and optimization. Infuluened by the noises from

external environment and electromagnetic interference during the
LiDAR is an active imaging technology, which calculates the position waveform acquisition process, it is necessary to filter and smooth the
and related parameters of ground objects by transmitting laser pulses original data to improve the signal-to-noise ratio. Hofton [7] used
and receiving the echo reflected by the target. Owing to the advantages Gaussian smoothing filter, Li used wavelet denoising [8], Ma used PTD
of small beam divergence, high peak power and good directionality [1], (Progressive TIN densification) filter [9], Zhang used EMD (Empirical
it is an important research field of detecting complex ground objects Mode Decomposition) filter [10], Wang used mean filter to filter the
with LiDAR. LiDAR has important application value in marine biological original waveform [11], Chen used EEMD-SVD-LWT to denoising LiDAR
detection [2], ground object classification [3], forestry monitoring [4], signal [12], and Xu proposed a two-level denoising framework with
low-level wind shear intensity test [5], retrieval of atmospheric aerosol singular value decomposition and adaptive wavelet denoising [13].
[6], etc. Traditional airborne lidar records limited number of echoes and Since the multi-target superimposed echo is considered to be the su­
provides information such as intensity and echo times, while the full- perposition of a series of basic waveforms, a single waveform model
waveform airborne LiDAR can record the complete echo waveform at should be selected as the decomposition model. In terms of decompo­
short-timeintervals. Meanwhile, a variety of reflection information such sition model selection, Wagne used Gaussian function as the waveform
as amplitude, wave width, energy from the original data can be decomposition basis function [14]. For the case that the transmitted
extracted with higher ranging accuracy and surface resolution. pulse is not perfectly symmetrical, Chauve employed the generalized
By using the full-waveform LiDAR echo decomposition algorithm, Gaussian function [15] and lognormal function [16] as the decomposi­
obstacles and targets can be distinguished. Waveform decomposition is tion basis function. Shen considered partial normal distribution function
mainly composed of signal denoising, decomposition model selection, as the decomposition basis function [17]. Gu proposed multi-Gaussian

* Corresponding author.
E-mail address: [email protected] (X. Xu).

https://doi.org/10.1016/j.infrared.2023.104613
Received 16 June 2022; Received in revised form 3 December 2022; Accepted 18 February 2023
Available online 23 February 2023
1350-4495/© 2023 Elsevier B.V. All rights reserved.
X. Xu et al. Infrared Physics and Technology 130 (2023) 104613

Fig. 1. Waveform decomposition process.

model as the decomposition basis function [18]. After choosing the (1) LSTM (Long Short-Term Memory) neural network is proposed to
decomposition basis function, the next step is to judge the number of determine the number of Gaussian components, which is more
basis functions in the echo waveform and obtain initial evaluation pa­ accurate than inflection point method and deconvolution
rameters. Methods most adopted are deconvolution method [19] and method.
inflection point method [20]. The last step is to use an optimization (2) SSA (Sparrow Search Algorithm) is used to fit the denoised
algorithm to fit the denoised waveform signal to extract the parameters. waveform. The combination of LSTM and SSA algorithms breaks
Wang used DSC (differential cuckoo search algorithm) for parameter through the distance limit of full waveform echo decomposition
optimization [21]. Li used the improved EM (Expectation Maximum) and can improve the decomposition accuracy.
algorithm to optimize the parameters of the full-waveform LiDAR data
in forest areas [22]. Xu proposed the improved LM (Levenberg-Mar­ The remainder of this paper is organized as follows. In Section 2, the
quardt) algorithm, which accelerates the convergence speed and itera­ detail of the proposed method is described, including wavelet smooth­
tion efficiency [23]. Shen used MDE (Modified Differential Evolution) ing, LSTM network recognizing number of Gaussian components in
combined with chaotic algorithm to solve the optimal parameters [24]. waveform, and SSA parameter estimation. In Section 3, the construction
Lai used L-BFGS (limited-memory BFGS) algorithm to optimize the of simulation data sets for LSTM neural network and the training of
waveform parameters [25]. However, most of these intelligent optimi­ LSTM network are discribed. In Section 4, experiment are carried out to
zation algorithms are sensitive to the initial values of model parameters compare the performance of the proposed method with IP method, RL
and easy to fall into local optimal values, which will affect the fitting and MGD method. Finally conclusions and future prospects are given in
accuracy. Therefore, some scholars used swarm intelligence algorithm Section 5.
to solve the optimal waveform parameters [26–28]. Oliveira used GA
(Genetic Algorithm) [29], Liuses used PSO (Particle Swarm Optimiza­ 2. The proposed algorithm
tion) [30], and Pandurouses used DE (Differential Evolution) algorithm
to find the global waveform parameters [31], where all achieved great The flow chart of the full-waveform decomposition algorithm is
fitting effect. shown in Fig. 1.
In summary, most of intelligent optimization algorithms are easy to Step 1: Determine whether there is an effective waveform. Since the
fall into local optimal values based on inaccurate initial values of model front and end of the measured waveform is all noise, the standard de­
parameters. Researchers generally use swarm intelligence optimization viation of the front several points of the waveform is taken as the noise
algorithms to improve the accuracy of the waveform decomposition, level δ. Whether there is an echo signal is judged by the peak value of the
which is insensitive to the initial value. Moreover, the inflection point sampling point. When the peak value is lower than 3δ, the echo signal
method and deconvolution method is mainly used in parameter esti­ processing ends. Otherwise, it is considered that there is an effective
mation. However, the accuracy of the decomposition results using these waveform. Go to step 2.
two methods is poor when the targets are close. Step 2: Peak point of the waveform and sample points on the left and
The contribution of this paper is: right of the peak value are taken. These points are fed into the trained
LSTM network to extract the number of Gaussian components.

2
X. Xu et al. Infrared Physics and Technology 130 (2023) 104613

maximum point and inflection point (first-order zero crossing and

second-order zero crossing). Then, the estimated Gaussian component is
subtracted from the denoising waveform. Furthermore, the above
operation is performed on the remaining waveform until the local
maximum in the remaining waveform is less than the noise threshold or
there is no inflection point around the local maximum. The principle of
RL [19] method is that a received waveform is the convolution of a
transmitted waveform and the effective target cross section. The over­
lapping sub-waveforms can be detected from the denoising waveforms
by Richardson-Lucy method, and the initial estimated values of the
parameters of the sub waveform decomposition model can be obtained.
The principle of MGD [18] is that multi-Gaussian model are used as basic
component in decomposition, which are denoted as MGD-X (X repre­
sents the number of Gaussian components in multi-Gaussian model).
Fig. 2. Structure diagram of LSTM unit. However, the accuracy of the IP method is limited when the distance
between the target components in the echo waveform is close, and the
Afterwards, the original echo waveform is preprocessed by wavelet accuracy of the initial estimation of the decomposition model parame­
denoising. ters obtained by RL and MGD method is low, which will be proved by the
Step 3: When the number of Gaussian components in waveform is experimental results below. Therefore, LSTM neural network is trained
greater than 1, SSA is used to fit the multi-target echo waveform to to extract the number of Gaussian components in full waveform echo
obtain the parameters of each Gaussian component. Otherwise, LM is signal. LSTM is a special recurrent neural network, which can learn long-
used for waveform fitting to obtain parameters of echo waveform. term dependence, especially in long-term sequence prediction problems.
The standard LSTM is composed of forget gate, input gate, output gate
2.1. Wavelet denoising preprocessing and update gate. The unit structure of LSTM is shown in Fig. 2. The
storage and update of the previous information for time series by LSTM
In this paper, the wavelet function is used to filter the echo signal. In are determined by these gates. Gate is realized by Sigmoid function and
this method, the “dB4′′ wavelet is used to decompose the echo waveform point multiplication operation. In essence, it is a full connection layer.
by three-level detail coefficients and approximate coefficients, and the Its general formula can be discripted by equation (3):
”rigrsure“ threshold and ”soft“ method are used for thresholding. 1
Finally, the approximate coefficients and detail coefficients are used to g(x) = σ (Wx + b)σ(x) = x
(3)
1 + e−
reconstruct the pulse laser echo signal.
where, σ (x) is Sigmoid function, which can map the calculation results to
2.2. Selection of waveform decomposition model the interval of [0,1]. When it is 0, it means that no information can pass
through. When it is 1, it means that all information can pass through. Ct
The full-waveform LiDAR signal is a function of the echo amplitude is the cell state at tth iteration, ht is the hidden layer state at the tth
changing over time. The decomposition algorithm deals the multi-target iteration. xt can be regarded as the vector form of the data input into
echo signal as the superposition of several function components, and LSTM, W is weight matrix, b is offset value. tanh is hyperbolic tangent
each function component is the echo waveform of single target. The function, and its output is between − 1 and 1. The calculation formulas of
purpose of waveform decomposition is to determine the number of each gate in Fig. 2 are shown in Eqs. (4)-(9).
( )
function components and calculate the parameters of each function ft = σ Wf ⋅[ht− 1 , xt ] + bf (4)
component. In this paper, Gaussian function will be used as the basic
function of waveform decomposition, and the echo equation is shown as: it = σ (Wi ⋅[ht− 1 , xt ] + bi ) (5)
ϕ
∑ ϕ
∑
y= f (t) + ε = αi e − (t− μi )2 /2η2i
+ε (1) ̃ t = tanh(Wc ⋅[ht− 1 , xt ] + bc )
C (6)
i=1 i=1

Ct = ft *Ct− 1 + it *C
̃t (7)
where, f(t) is Gaussian function, φ is number of Gaussian functions, t is
time, α is amplitude of Gaussian function, μ is peak time of Gaussian ot = σ (Wo ⋅[ht− 1 , xt ] + bo ) (8)
function, η is standard deviation of Gaussian function, ε is the noise.
The distance between the ith target and the i + 1th target is: ht = ot *tanh(Ct ) (9)
L = |μi − μi+1 |*c/2 (2)
2.4. Accurate estimation based on SSA
where, c is velocity of light. The purpose of full-waveform decomposi­
tion algorithm is to determine φ, and then calculate three parameters of Sparrows are gregarious birds. In contrast with a lot of other birds,
each Gaussian function. sparrows are highly intelligent and have good memories. Among the
foraging sparrows, the identities of the sparrows are mainly divided into
two categories: producer and scrounger. The producers actively search
2.3. Parameter pre-estimation
for the food source, while the scroungers obtain food by the producers.
Furthermore, the evidence shows that the birds use behavioural strate­
In order to reduce the sensitivity of the full waveform echo decom­
gies flexibly, and switch between producing and scrounging. According
position algorithm to initial values, the swarm intelligence optimization
to these description of the sparrows, we can establish the mathematical
algorithm is used by merely knowing the number of single waveform.
model of the sparrow search algorithm. Firstly, according to the ideal
The most commonly used method is inflection point (IP) method or
shape of sparrows, the foraging rules of sparrows are set [32]:
Richardson-Lucy (RL) deconvolution method. The principle of IP [7]
method is that for each Gaussian component, the value of three pa­
rameters of Gaussian component can be estimated according to the local

3
X. Xu et al. Infrared Physics and Technology 130 (2023) 104613

(4) The sparrows with the higher energy would act as the producers.
Several starving scroungers are more likely to fly to other places
for food in order to gain more energy.
(5) The scroungers follow the producer who can provide the best
food to search for food. In the meantime, some scroungers may
constantly monitor the producers and compete for food in order
to increase their own predation rate.
(6) The sparrows at the edge of the group quickly move toward the
safe area to get a better position when aware of danger, while the
sparrows in the middle of the group randomly walk to be closer to
others.

In mathematics, we use virtual sparrows to find food. The position of

sparrows can be represented as:
⎡ ⎤
x1,1 x1,2 ⋯ ⋯ x1,d
⎢ x2,1 x2,2 ⋯ ⋯ x1,d ⎥
X=⎢ ⎣ ⋮
⎥ (10)
⋮ ⋮ ⋮ ⋮ ⎦
xn,1 xn,2 ⋯ ⋯ xn,d
Fig. 3. Normalization of simulated waveform and real echo.
where n is the number of sparrows, d is the dimension of the variable to
(1) The producers typically have high levels of energy reserves and be optimized. Then, the fitness values of all sparrows can be expressed
provide foraging areas or directions for all scroungers. It is as:
responsible for identifying areas where rich food sources can be ⎡ ⎤
f ([x1,1 x1,2 ⋯ ⋯ x1,d ])
found. The level of energy reserves depends on the assessment of ⎢ f ([x2,1 x2,2 ⋯ ⋯ x1,d ]) ⎥
the fitness values of the individuals. FX = ⎢
⎣ ⋮
⎥ (11)
⋮ ⋮ ⋮ ⋮ ⎦
(2) Once the sparrow detects the predator, the individuals begin to f ([xn,1 xn,2 ⋯ ⋯ xn,d ])
chirp as alarming signals. When the alarm value is greater than
the safety threshold, the producers need to lead all scroungers to where the value of each row in FX represents the fitness value of the
the safe area. individual. The data processed is dual waveform echo superposition
(3) Each sparrow can become a producer as long as it finds better waveform, so the fitness value FXi in line i is shown in Eq. (12):
food sources, but the ratio of the producers and the scroungers is
unchanged in the whole population.

Fig. 4. Simulation waveform under differernt kinds of noise. (a) simulation waveform of the original echo (b) the waveform after adding Gaussian white noise (c) the
waveform after adding Poisson noise (d) the waveform after adding Uniform noise.

4
X. Xu et al. Infrared Physics and Technology 130 (2023) 104613

Fig. 5. Construction of LSTM network training simulation data sets.

Fig. 6. LSTM network training results with Gaussian white noise. (a)Accuracy rate curve (b) Loss curve.

Fig. 7. LSTM network training results with Poisson noise. (a) Accuracy rate curve (b) Loss curve.

5
X. Xu et al. Infrared Physics and Technology 130 (2023) 104613

Fig. 8. LSTM network training results with Uniform noise. (a) Accuracy rate curve (b) Loss curve.

Fig. 9. Full waveform echo acquisition experimental platform.

Table 1
System structural parameters of the experimental LiDAR.
System Parameter Value

Laser Wavelength λ 1064 nm

Pulse peak power 75 W
Pulse width (FWHM) 6 ns
Repetition frequency 10 kHz
Oscilloscope Sampling rate 2.5 GS/s
Computer GPU RTX2070
CPU I5-10400
RAM 16G

responsible for searching food and guiding the movement of the entire
population, the producers can search for food in a broader range of the
places than that of the scroungers. According to rules (1) and (2), during
Fig. 10. Mechanism of receiving echo. each iteration, the location of the producer is updated as:
⎧ ( )
⎛ 2 2
⎞2 ⎪ − i
− (z− xi,2 ) − (z− xi,5 ) ⎪ Xi,jt ⋅exp
⎪ if R2 < ST
∑
501
⎜ ⎟ ⎨ α⋅itermax
⎝xi,1 *e i,3 + xi,4 *e i,6 − P(z) ⎠ (12)
2 2
(13)
2*x 2*x
FXi = Xi,jt+1 = ( )
z=1 ⎪
⎪ t Ub − Lb
⎪
⎩ Xi,j + Q⋅ if R2 ≥ ST
n
where z is 1 to 501, because the oscilloscope used in the experiment
t
collects data at 501 sampling points each time. P(z) represents the where, t is number of current iterations, Xi,j represents the value of the
waveform after denoising. The smaller the FXi is, the better the adapt­ jth dimension of the ith sparrow at iteration t, itermax is a constant with
ability, and the higher the waveform fitting degree are. the largest number of iterations, R2 is a random value from 0 to 1, which
In the SSA, the producers with better fitness values have the priority represents the alarm value, ST is a value from 0.5 to 1, which represents
to obtain food in the search process. In addition, Since the producers are safety threshold. Q is a random number which obeys normal

6
X. Xu et al. Infrared Physics and Technology 130 (2023) 104613

Fig. 11. Start single processing (a) Original waveform (b) Denoised waveform (c) waveform fitting results after LM algorithm.

t
distribution. Ub and Lb is the upper and lower limit ofXi,j . ⎧ ⃒ ⃒
⃒ t ⃒
If R2 is less than ST, there are no predators around, and the producer ⎪
⎪
⎪ XPt + R⋅⃒Xi,jt − Xbest ⃒ if fi > fg
⎨
enters the wide area search mode. Otherwise, it means that some spar­ Xi,jt+1 =
( )
(15)
rows have discovered the predator, and all sparrows need to fly quickly ⎪
⎪ M⋅ Xi,jt − Xworse
t
⎪ Xt +
⎩ if fi = fg
to other safe areas. i,j
(fi − fw + ε)
As for the scroungers, they need to satisfy the rules (4) and (5). As
t t
mentioned above, some scroungers monitor the producers more where Xbest is the current global optimal location. Xworse is the current
frequently. Once they find that the producer has found great food, they global worst location. R is a random number which obeys normal dis­
immediately leave their current position to compete for food. If they tribution with mean value of 0 and standard deviation of 6. M is a
win, they can get the food of the producer immediately, otherwise, they random value from − 1 to 1. fi is fitness value of current sparrow. fg and
continue to execute the rules (5). The position update formula for the fw is current global optimal and worst fitness values. ε is the smallest
scrounger is described as: constant so as to avoid zero-division-error. If fi is larger than fg, which
⎧ represents the sparrow is at the edge of the group and makes the spar­
⎨ Lb + (Ub − Lb)*exp( − |W| ) if i > N/2
⎪ rows who feel dangerous fly to the central groupXbest t
. fi = fg shows that
( )
Xi,jt+1 = Ub − Lb (14) the sparrows, which are in the middle of the population, are aware of the
⎪
⎩ XPt + E⋅ otherwise
n danger and need to move closer to the others.

where XPt is the optimal position occupied by the producer. W is a 3. LSTM network training
random number which obeys normal distribution with mean value of
0 and standard deviation of 5. E is a random number, which obeys The training data set used in this paper is obtained by simulation
normal distribution with mean value of 0 and standard deviation of 0.1. data. It is known that the full waveform lidar echo equation of N targets
If i is larger than n/2, it suggests that the ith scrounger with the worse is [33]:
fitness value is most likely to be starving. In addition, we assume that ( )(
these sparrows, aware of the danger, account for 20% of the total pop­ ∑N
πP0 D2 η2atm ηsys fr (ψ k )cosψ k ξ2k b2k √̅̅̅̅̅
Pr (t) = √ ̅̅̅̅̅̅̅ exp − + erf ( ak (Bk
ulation. According to rules (6), the mathematical model can be k=1
4ωk R2rk 2ak c2 τ2 4ak
expressed as: bk √̅̅̅̅̅ bk
)
+ )) − erf ( ak (Ak + ))
2ak 2ak
(16)

7
X. Xu et al. Infrared Physics and Technology 130 (2023) 104613

Fig. 12. Dual target echo processing process (a) Original waveform (b) Denoised waveform (c) waveform fitting results using SSA (d) Iteration procedure of SSA.

2c2 τ2 +4ω2k tan2 ψ k with Gaussian white noise, Poisson noise, and Uniform noise are shown
where, ξk = 2Rrk − ct.ak = c2 τ2 ω2k
.bk = 4ξck2tan
τ2 .P0 is peak power of
ψk
in Fig. 6, Fig. 7, and Fig. 8, respectively. In Fig. 6, the training accuracy
the pulsed laser.τ is laser output pulse width.ωk is laser spot diameter. Ak rate reached 100% in the 181th iteration, the validation accuracy rate
and Bk is boundary value of illumination area of light spot on target.Rrk reached 100% in the 183th iteration, the training loss value finally
is distance from target to laser.ψ k is the angle of the irradiated target decreased to 0.00316, and the validation loss value finally decreased to
relative to the normal plane of the laser beam. Other parameters are 0.00229. In Fig. 7, the training accuracy rate reached 100% in the 178th
fixed values and affect the amplitude of the waveform. The single iteration, the validation accuracy rate reached 100% in the 210th iter­
simulation waveform can be compared with the experimental waveform ation, the training loss value finally decreased to 0.00147, and the
for normalization as shown in Fig. 3. In this paper, 10,000 sets of validation loss value finally decreased to 0.00248. In Fig. 8, the training
simulated waveforms are used for training. In addition, in order to accuracy rate reached 99% in the 192th iteration, the validation accu­
ensure that the simulated waveform is close to real echo, Gaussian white racy rate reached 99% in the 218th iteration, the training loss value
noise, Poisson noise and Uniform noise are added to the simulated finally decreased to 0.00161, and the validation loss value finally
waveform, as shown in Fig. 4. decreased to 0.00242.
10,000 sets of data are divided into 10 groups, and each group has
1000 sets, which corresponds to one situation that targets are irradiated 4. Experiments
by the laser. As shown in Fig. 5, the first nine groups are the echoes of
two targets. From the first group to the ninth group, the area of target 1 4.1. Construction of experimental platform
is increasing, and the area of target 2 is decreasing. In each group of
1000 sets data, the distance between target 1 and the laser remains After LSTM network training, dual echo waveform and single echo
unchanged, and the distance between target 2 and target 1 continues to waveform are collected to verify the method proposed in this paper. The
increase. The distance between target 1 and target 2 increases from 0.4 experimental setup is shown in Fig. 9. The mechanism of receiving echo
m and the interval is 0.001 m. The tenth group of data is single echo is shown in Fig. 10. The laser is emitted and divided into two beams after
waveform. The distance between the target and the laser is increasing. passing through the light splitting prism. One beam is directly received
The initial distance is 10 m and the interval is 0.001 m. Then 10,000 sets by the detector 2 as the starting signal, and the other beam irradiates the
of simulated data are randomly divided into two groups and fed into target. Both targets are mounted on the sliding block, where the distance
LSTM network for training, where 7000 sets for training and 3000 sets between can be adjusted by the sliding rail. The key system parameters
for validation. The success rate and loss curve of the training process are listed in Table 1. In the single waveform echo experiment, ten sets of

8
X. Xu et al. Infrared Physics and Technology 130 (2023) 104613

Fig. 13. Forecast results under LSTM networks trained with three kinds of background noises (a) LSTM networks trained with Gaussian white noise (b) LSTM
networks trained with Poisson noise (c) LSTM networks trained with Uniform noise.

Table 2
Decomposition accuracy of six methods.
D1(m) D2(m) Decomposition accuracy

Prop IP RL MGD-2 MGD-3 MGD-4

10 10.4 100% 0% 68% 72% 82% 78%

10 10.45 100% 0% 96% 92% 84% 78%
10 10.5 100% 0% 80% 92% 70% 60%
10 10.55 100% 0% 96% 90% 78% 82%
10 10.6 100% 8% 98% 98% 84% 84%
10 10.65 100% 34% 98% 98% 84% 84%
10 10.7 100% 72% 100% 98% 86% 82%
10 10.75 100% 92% 100% 94% 82% 82%
10 10.8 100% 98% 100% 100% 88% 86%
10 10.85 100% 100% 98% 90% 82% 82%
10 10.9 100% 100% 98% 94% 88% 88%
10 10.95 100% 88% 98% 94% 84% 80%
10 11 100% 96% 98% 90% 74% 74%

4.2. The echo decomposition process of the algorithm proposed in this

paper
Fig. 14. The ground distance and peak distance of dual waveform echo.
The following figures show the decomposition process of the
echoes with the distance of 10 m, 10.2 m, 10.4 m, 10.6 m, 10.8 m and 11 experimental waveform. In order to reduce the calculation time, take the
m were measured respectively. In the dual waveform echo experiment, peak point and both 40 points on the left and right side of the peak point,
the distance of target 1 D1 is unchanged at 10 m, and the distance of the a total of 81 points as the original waveform. Fig. 11 (a) is the original
target 2 D2 changed at 10.4 m, 10.45 m, 10.5 m, 10.55 m, 10.6 m, 10.65 waveform of start single. The original waveform is denoised by wavelet,
m, 10.7 m, 10.75 m, 10.8 m, 10.85 m, 10.9 m, 10.95 m and 11 m. 50 sets and the denoised waveform is shown in Fig. 11 (b). The noise level of the
of dual waveform echo data were measured in each position waveform in Fig. 11 (b) δ = 0.0146 V, while the noise level of waveform
respectively. in Fig. 11 (a) δ = 0.0280 V. Then, the denoised waveform is fitted by LM
algorithm, the fitting result is shown in Fig. 11 (c), where α1 =
0.228、μ1 = 6.7600 × 10-8、η1 = 11.0325.
Fig. 12 (a) shows the experimental original echo waveform when
laser irradiates on two targets, in which the actual distance between the

9
X. Xu et al. Infrared Physics and Technology 130 (2023) 104613

Table 3 Table 6
Decomposition result of six methods. Decomposition success ratepeak result of six methods.
D1(m) D2(m) RMSEground(cm) D1(m) D2(m) Decomposition success ratepeak

Prop IP RL MGD-2 MGD-3 MGD-4 Prop IP RL MGD-2 MGD-3 MGD-4

10 10.4 12.16 16.21 32.50 15.79 22.19 22.02 10 10.4 48% 0% 12% 32% 38% 40%
10 10.45 10.44 19.01 32.54 17.39 20.77 21.14 10 10.45 72% 0% 28% 48% 44% 38%
10 10.5 12.82 15.31 25.74 14.12 18.43 34.06 10 10.5 76% 0% 56% 64% 58% 60%
10 10.55 15.68 22.49 26.53 18.61 34.46 33.80 10 10.55 88% 0% 78% 76% 78% 68%
10 10.6 11.23 59.05 13.35 13.53 35.27 35.38 10 10.6 88% 8% 88% 78% 80% 78%
10 10.65 9.71 67.63 13.51 12.61 17.54 16.50 10 10.65 88% 34% 90% 86% 82% 82%
10 10.7 9.75 10.08 12.45 13.60 13.90 14.12 10 10.7 92% 70% 98% 98% 88% 82%
10 10.75 9.16 9.23 9.19 11.27 30.20 13.52 10 10.75 96% 88% 100% 92% 82% 84%
10 10.8 5.65 6.14 9.11 12.51 12.68 16.28 10 10.8 100% 98% 100% 100% 94% 92%
10 10.85 6.24 6.44 15.95 17.30 15.01 14.46 10 10.85 100% 100% 98% 90% 82% 82%
10 10.9 6.83 7.12 9.91 13.87 15.70 12.22 10 10.9 94% 92% 94% 94% 88% 90%
10 10.95 4.24 4.24 10.15 12.89 17.45 16.72 10 10.95 100% 88% 94% 94% 84% 80%
10 11 2.98 3.18 10.74 15.70 31.39 17.62 10 11 98% 94% 98% 90% 78% 78%

iteration.
Table 4
Decomposition result of six methods.
4.3. Comparison between LSTM networks under different kinds of noise
D1(m) D2(m) RMSEpeak(cm)

Prop IP RL MGD-2 MGD-3 MGD-4 All 650 sets of dual experimental waveform echoes and 60 sets of
10 10.4 12.87 40 30.27 27.14 20.20 21.48 single experimental waveform echoes are fed into the LSTM networks
10 10.45 10.41 45 26.47 15.65 20.68 24.23 trained with three kinds of background noises, respectively. As shown in
10 10.5 13.89 50 24.99 22.14 23.82 38.37 Fig. 13, the predict accuracy rates of LSTM networks trained with
10 10.55 5.49 55 17.18 18.21 34.11 33.28
Gaussian white noise, Poisson noise and Uniform noise are 100%,
10 10.6 6.48 57.59 11.33 8.45 32.11 32.55
10 10.65 6.58 52.86 28.20 28.85 31.02 30.73
99.01% and 95.77%, respectively. The predict accuracy rates of Uniform
10 10.7 5.14 37.26 5.55 5.95 8.62 9.14 noise is lowest. For laser system, Gaussian white noise and Poisson noise
10 10.75 4.79 21.64 4.99 12.55 29.27 15.23 are more suitable as the background noise. Both of predict accuracy
10 10.8 3.19 11.78 3.58 3.49 6.85 9.88 rates are higher than 99%.
10 10.85 3.40 3.42 13.09 21.21 24.3 25.19
10 10.9 4.20 4.39 10.82 11.8 13.95 12.27
10 10.95 3.04 33.03 10.82 10.55 14.72 15.12 4.4. Comparison of result between conventional algorithm and SSA
10 11 3.67 20.35 9.52 12.39 28.16 15.16

All 650 sets of dual experimental waveform echoes are decomposed

using Prop (proposed method), IP, RL, MGD-2, MGD-3 and MGD-4. The
Table 5 accuracy of these methods to distinguish the number of echo waveform
Decomposition success rateground of six methods. components is defined as the decomposition accuracy. The RMSE of
D1(m) D2(m) Decomposition success rateground
ground distance (RMSEground) and RMSE of peak distance (RMSEpeak)
are used to verify the accuracy of these methods. Fig. 14 shows the
Prop IP RL MGD-2 MGD-3 MGD-4
ground distance and peak distance in a dual waveform echo. The ground
10 10.4 58% 8% 8% 40% 36% 34% distance is calculated by the time difference between the location of the
10 10.45 54% 6% 6% 48% 48% 42%
start signal and that of the last echo component in fitting echo wave­
10 10.5 58% 30% 30% 52% 36% 34%
10 10.55 52% 36% 36% 50% 50% 48% form. The peak distance is calculated by the time difference between the
10 10.6 56% 42% 42% 54% 44% 42% location of the first echo component and that of the last echo component
10 10.65 74% 52% 52% 64% 54% 58% in fitting echo waveform. The decomposition accuracies, the RMSE
10 10.7 76% 62% 62% 52% 46% 48%
ground, and the RMSEpeak of dual experimental waveform echoes using
10 10.75 72% 72% 72% 66% 56% 68%
Prop, IP, RL, MGD-2, MGD-3 and MGD-4 are listed in Table 2, Table 3,
10 10.8 92% 64% 64% 56% 48% 48%
10 10.85 92% 62% 62% 72% 66% 70% and Table 4, respectively. The decomposition accuracies of the Prop are
10 10.9 90% 74% 74% 76% 72% 76% 100% in all distances of D2. However, the decomposition accuracies of
10 10.95 98% 92% 92% 80% 76% 76% the IP method are less than 50% when D2 is less than 10.7 m. Moreover,
10 11 100% 98% 98% 70% 66% 64%
the decomposition accuracies of the RL method are larger than 90% in
all distances of D2 except 10.4 m and 10.5 m, the decomposition accu­
two targets Lreal = 0.7 m. The original waveform is fed into the LSTM racies of the MGD-2 method are larger than 90% in all distances of D2
network to obtain the estimated number of waveforms, and the number except 10.4 m, and the decomposition accuracies of the MGD-3 and
of Gaussian components is 2. Then, the original waveform is denoised by MGD-4 method are less than 90% in all distances of D2.The RMSEsground
wavelet, and the denoised waveform is shown in Fig. 12 (b). The noise of the Prop method are less than 10 cm when D2 is larger than 10.6 m,
level of the waveform in Fig. 12 (a) is 0.0114 V, while the noise level of and the RMSEspeak of the Prop method are less than 10 cm when D2 is
waveform in Fig. 12 (b) is 0.0065 V. Finally, SSA is used to fit the larger than 10.5 m. The RMSEspeak of RL method and IP method are close
denoised waveform. In SSA, the number of sparrows n = 1000, and the to those of the Prop method when decomposition accuracies are 100%.
fitting result is shown in Fig. 12 (c), where α1 = 0.1957, μ1 = 1.5865 × Although RL method can distinguish more dual waveforms than IP
10-10, η1 = 4.7657, α2 = 0.1745, μ2 = 2.0674 × 10-10, η2 = 5.4718. The method, the RMSEsground of RL method are larger than that of IP method
calculated distance between two targets Lcal = 0.7214 m. In SSA the in all distances except 10.6 m, 10.65 m and 10.75 m. The RMSEsground of
decline procedure of fitness value in the iterative process is shown in MGD-2 are less than those of RL method when D2 less than 10.7 m. The
Fig. 12 (d), the fitness value FX decreased to 27.8052 after the 27th RMSEspesk of MGD-2 are less than those of RL method when D2 is from
10.4 m to 10.95 m. The RMSEground and RMSEpeak of MGD-3 and MGD-4

10
X. Xu et al. Infrared Physics and Technology 130 (2023) 104613

Fig. 15. Decomposition results of dual waveform echo at different distances D2. (a) D2 = 10.9 m (b) D2 = 10.8 m (c) D2 = 10.7 m (d) D2 = 10.6 m (e) D2 = 10.5 m (f)
D2 = 10.4 m.

are larger than MGD-2 in most distances of D2. It means MGD method ground distance and peak distance using the Prop method are all larger
can get more accurate results when distance between targets is small, than IP, RL, MGD-2, MGD-3 and MGD-4 method.
but when distance between targets is large, the result might be worse. At last, Fig. 15 shows the waveform decomposition results with
For the waveforms collected by the detector in this experiment multi- different distances of D2, and Fig. 16 shows the decline procedure of
Gaussian model do not get significantly better results than Gaussian fitness value in SSA iteration. In the iteration process of SSA with each
model. The RMSEground and RMSEpeak of the Prop are no larger than distance D2, the fitness value all reaches the best before 60 iterations.
those of the IP, RL, MGD-2, MGD-3 and MGD-4 method in all distances.
In conclusion, the proposed method has stronger stability in decompo­ 5. Conclusions
sition accuracy than IP, RL, MGD-2, MGD-3 and MGD-4 method.
All 650 sets of dual waveform echo data are processed by six This paper presents a method of waveform decomposition using
methods. The decomposition results whose error less than 0.1 m are LSTM neural network combined with SSA. The network is trained
recorded as successful decompositions. The ground distance success rate through the simulation data sets under different kinds of background
of the conventional methods and the method proposed are shown in noise, and the trained network is used to preprocess the experimental
Table 5. The peak distance success rate of the conventional methods and data. Compared with the infection point method and Richardson-Lucy
the method proposed are shown in Table 6. When the difference be­ deconvolution method, LSTM network has higher accuracy when
tween D1 and D2 is less than 0.7 m. The decomposition success rate of judging the number of Gaussian components in echoes. Because the

11
X. Xu et al. Infrared Physics and Technology 130 (2023) 104613

Fig. 16. Iteration process of dual waveform echo at different distances D2. (a) D2 = 10.9 m (b) D2 = 10.8 m (c) D2 = 10.7 m (d) D2 = 10.6 m (e) D2 = 10.5 m (f) D2 =
10.4 m.

LSTM network cannot extract the pre-estimated value of Gaussian 0.25 m less than that of IP method, 0.1 m less than that of RL and
component parameters which can be obtained by the inflection point MGD method, which means multi-waveform echo with smaller
method, the SSA is used in waveform decomposition and compared with distance can be successfully recognized and decomposed.
conventional algorithms. The following conclusions are drawn:
Declaration of Competing Interest
(1) Under the same data set, the proposed method has the smaller
RMSE of ground distance and peak distance than IP, RL and MGD The authors declare that they have no known competing financial
method. interests or personal relationships that could have appeared to influence
(2) The proposed method breaks through the distance limit of the work reported in this paper.
waveform decomposition. When the decomposition success rate
of ground distance is greater than 70%, the minimum decompo­ Data availability
sition distance of the proposed method is 0.1 m less than that of IP
and RL method, 0.2 m less than that of MGD method. When the Data will be made available on request.
decomposition success rate of peak distance is greater than 70%,
the minimum decomposition distance of the proposed method is

12
X. Xu et al. Infrared Physics and Technology 130 (2023) 104613

Acknowledgments [16] A. Chauve, C. Mallet, F. Bretar, S. Durrieu, M.P. Deseilligny, W. Puech, Processing
full-waveform Lidar data: modelling raw signals, Int. Arch. Photogramm. Remote
Sens. Spat. Inf. Sci. 2007 (2004) 2007.
This research is funded by National Natural Science Foundation of [17] S. Jun, S. Jianhua, S. Jiatong, H. Yan, Echo characteristics of multiplane targets
China (Grant No. 51805146), Changzhou Sci&Tech Program (Grant No. detection using pulsed laser, Chinese J. Lasers 45 (11) (2018) 1110004, https://
CE20215041), and the Fundamental Research Funds for the Central doi.org/10.3788/CJL201845.1110004.
[18] Z. Gu, J. Lai, C. Wang, W. Yan, Y. Ji, Z. Li, Decomposition of LiDAR waveforms
Universities (Grant No. B220202023). with negative tails by Gaussian mixture model, Opt. Eng. (2021) 60, https://doi.
org/10.1117/1.oe.60.5.054102.
References [19] T. Zhou, S.C. Popescu, Bayesian decomposition of full waveform LiDAR data with
uncertainty analysis, Remote Sens. Environ. (2017) 200, https://doi.org/10.1016/
j.rse.2017.08.012.
[1] Y. Wang, H. Weinacker, B. Koch, A Lidar point cloud based procedure for vertical [20] D. Yihao, Z. Aiwu, L. Zhao, W. Shumin, W. Jingmeng, Y. Qiuhong, A Gaussian
canopy structure analysis and 3D single tree modelling in forest, Sensors (2008) 8, inflexion points matching method for gaussian decomposition of airborne LiDAR
https://doi.org/10.3390/s8063938. waveform data, Laser Optoelectron. Prog. 51 (10) (2014), 102801, https://doi.org/
[2] P. Chen, D. Pan, Ocean optical profiling in South China sea using airborne LiDAR, 10.3788/LOP51.102801.
Remote Sens. 11 (2019) 1826, https://doi.org/10.3390/rs11151826. [21] M. Wang, S. Xiong, M. Chen, P. He, A waveform decomposition technique based on
[3] A. Schmidt, F. Rottensteiner, U. Sörgel, Water-land-classification in coastal areas wavelet function and differential cuckoo search algorithm, Soft Comput. (2021)
with full waveform lidar data, Photogramm. Fernerkundung, Geoinf. (2013; 25, https://doi.org/10.1007/s00500-021-05583-x.
2013.), https://doi.org/10.1127/1432-8364/2013/0159. [22] Q. Li, W. Zhou, C. Li, Use of airborne LiDAR to estimate forest stand characteristics,
[4] T. Allouis, S. Durrieu, P. Couteron, A new method for incorporating hillslope effects IOP Conf. Ser. Earth Environ. Sci. 17 (2014), https://doi.org/10.1088/1755-1315/
to improve canopy-height estimates from large-footprint LIDAR waveforms, IEEE 17/1/012252.
Geosci. Remote Sens. Lett. (2012) 9, https://doi.org/10.1109/ [23] F. Xu, F. Li, Y. Wang, Modified Levenberg-Marquardt-based optimization method
LGRS.2011.2179635. for LiDAR waveform decomposition, IEEE Geosci. Remote Sens. Lett. (2016) 13,
[5] H. Zhang, S. Wu, Q. Wang, B. Liu, B. Yin, X. Zhai, Airport low-level wind shear lidar https://doi.org/10.1109/LGRS.2016.2522387.
observation at Beijing Capital International Airport, Infrared Phys. Technol. (2019) [24] F.C. Chang, H.C. Huang, A refactoring method for cache-efficient swarm
96, https://doi.org/10.1016/j.infrared.2018.07.033. intelligence algorithms, Inf. Sci. (Ny) (2012) 192, https://doi.org/10.1016/j.
[6] Y. Hu, X. Dong, N. Zhao, L. Guo, X. Zhao, Fast retrieval of atmospheric CO2 ins.2010.02.025.
concentration based on a near-infrared all-fiber integrated path coherent [25] X.D. Lai, N.N. Qin, X.S. Han, J.H. Wang, W.G. Hou, Iterative decomposition
differential absorption lidar, Infrared Phys. Technol. (2018) 92, https://doi.org/ method for small foot-print LiDAR waveform, Hongwai Yu Haomibo Xuebao/J.
10.1016/j.infrared.2018.07.004. Infrared Millim. Waves (2013) 32, https://doi.org/10.3724/SP.
[7] M.A. Hofton, J.B. Minster, J.B. Blair, Decomposition of laser altimeter waveforms, J.1010.2013.00319.
IEEE Trans. Geosci. Remote Sens. (2000) 38, https://doi.org/10.1109/36.851780. [26] A. Slowik, H. Kwasnicka, Nature Inspired Methods and Their Industry
[8] D. Li, L. Xu, X. Xie, X. Li, J. Chen, J. Chen, Co-path full-waveform LiDAR for Applications-Swarm Intelligence Algorithms, IEEE Trans Ind Informatics (2018)
detection of multiple along-path objects, Opt. Lasers Eng. (2018) 111, https://doi. 14, https://doi.org/10.1109/TII.2017.2786782.
org/10.1016/j.optlaseng.2018.08.009. [27] A. Alomari, W. Phillips, N. Aslam, F. Comeau, Swarm intelligence optimization
[9] H. Ma, W. Zhou, L. Zhang, DEM refinement by low vegetation removal based on techniques for obstacle-avoidance mobility-assisted localization in wireless sensor
the combination of full waveform data and progressive TIN densification, ISPRS J. networks, IEEE Access. (2017) 6, https://doi.org/10.1109/ACCESS.2017.2787140.
Photogramm. Remote Sens. (2018) 146, https://doi.org/10.1016/j. [28] F.K.H. Phoa, A Swarm Intelligence Based (SIB) method for optimization in designs
isprsjprs.2018.09.009. of experiments, Nat. Comput. 16 (2017), https://doi.org/10.1007/s11047-016-
[10] Z. Zhang, X. Liu, R. Shu, F. Xie, F. Wang, Z. Liu, et al., A novel noise reduction 9555-4.
method for space-borne full waveforms based on empirical mode decomposition, [29] A.A.A. de Oliveira, J.A.S. Centeno, F.S. Hainosz, Point cloud generation from
Optik (Stuttg) (2020) 202, https://doi.org/10.1016/j.ijleo.2019.163581. gaussian decomposition of the waveform LASER signal with genetic algorithms,
[11] C. Wang, F. Tang, L. Li, G. Li, F. Cheng, X. Xi, Wavelet analysis for ICESat/GLAS Bol. Ciencias Geod. (2018;) 24, https://doi.org/10.1590/S1982-
waveform decomposition and its application in average tree height estimation, 21702018000200018.
IEEE Geosci. Remote Sens. Lett. (2013) 10, https://doi.org/10.1109/ [30] D. Li, L. Xu, X. Li, Full-waveform LiDAR echo decomposition based on wavelet
LGRS.2012.2194692. decomposition and particle swarm optimization, Meas. Sci. Technol. (2017) 28,
[12] X. Cheng, J. Mao, J. Li, et al., An EEMD-SVD-LWT algorithm for denoising a lidar https://doi.org/10.1088/1361-6501/aa5c1e.
signal, Measurement 168 (3) (2021), 108405. [31] M.A. Panduro, C.A. Brizuela, A comparative analysis of the performance of GA,
[13] X. Xu, M. Luo, Z. Tan, R. Pei, Echo signal extraction method of laser radar based on PSO and DE for circular antenna arrays, IEEE Antennas Propag. Soc. AP-S Int.
improved singular value decomposition and wavelet threshold denoising, Infrared Symp. (2009), https://doi.org/10.1109/APS.2009.5171514.
Phys. Technol. (2018) 92, https://doi.org/10.1016/j.infrared.2018.06.028. [32] J. Xue, B. Shen, A novel swarm intelligence optimization approach: sparrow search
[14] W. Wagner, A. Ullrich, V. Ducic, T. Melzer, N. Studnicka, Gaussian decomposition algorithm, Syst. Sci. Control Eng (2020) 8, https://doi.org/10.1080/
and calibration of a novel small-footprint full-waveform digitising airborne laser 21642583.2019.1708830.
scanner, ISPRS J. Photogramm. Remote Sens. (2006) 60, https://doi.org/10.1016/ [33] W. Jiali, X. Xiaobin, W. Jialin, T. Zhiying, L. Minzhou, Echo characteristics of
j.isprsjprs.2005.12.001. multiplane targets detection using pulsed laser, Chinese J. Lasers 49 (6) (2022)
[15] A. Chauve, C. Vega, S. Durrieu, F. Bretar, T. Allouis, M.P. Deseilligny, et al., 0604004, https://doi.org/10.3788/CJL202249.060400.
Advanced full-waveform lidar data echo detection: assessing quality of derived
terrain and tree height models in an alpine coniferous forest, Int. J. Remote Sens.
(2009) 30, https://doi.org/10.1080/01431160903023009.

13