Applied Soft Computing Journal xxx (xxxx) xxx

Contents lists available at ScienceDirect

Applied Soft Computing Journal

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

A gradient boosting decision tree based GPS signal reception

classification algorithm
∗
Rui Sun a,b,c , , Guanyu Wang a , Wenyu Zhang a , Li-Ta Hsu d , Washington Y. Ochieng e
a
College of Civil Aviation, Nanjing University of Aeronautics and Astronautics, Nanjing 210016, China
b
Shandong Provincial Key Laboratory of Optical Astronomy and Solar-Terrestrial Environment, Shandong University, Weihai 264209, China
c
State Key Laboratory of Geo-Information Engineering, Xi’an Research Institute of Surveying and Mapping, Xi’an 710054, China
d
Interdisciplinary Division of Aeronautical and Aviation Engineering, The Hong Kong Polytechnic University, Hong Kong
e
Center for Transport Studies, Imperial College London, London SW7 2AZ, UK

article info a b s t r a c t

Article history: In urban areas, GPS signals are often reflected or blocked by buildings, which causes multipath
Received 15 May 2019 effects and non-line-of-sight (NLOS) reception respectively consequently degrading GPS positioning
Received in revised form 14 October 2019 performance. While improved receiver design can reduce the effect of multipath to some extent, it
Accepted 15 November 2019
cannot deal with NLOS. Modelling methods based on measurements have shown promise to reduce
Available online xxxx
the effect of NLOS signal reception. However, this depends on their ability to accurately and reliably
Keywords: classify line-of-sight (LOS), multipath and NLOS signals. The traditional method is based on one feature
GPS using signal strength as measured by the carrier to noise ratio, C/N0 . However, this feature is ineffective
GBDT in capturing the characteristics of multipath and NLOS in all environments. In this paper, to improve
Urban canyon the accuracy of signal reception classification, we are using the three features of C/N0 , pseudorange
Multipath
residuals and satellite elevation angle with a gradient boosting decision tree (GBDT) based classification
NLOS
algorithm. Experiments are carried out to compare the proposed algorithm with classifiers based on
decision tree, distance weighted k-nearest neighbour (KNN) and the adaptive network-based fuzzy
inference system (ANFIS). Test results from static receivers in urban environments, show that the GBDT
based algorithm achieves a classification accuracy of 100%, 82% and 86% for LOS, multipath and NLOS
signals, respectively. This is superior to the other three algorithms with the corresponding results of
100%, 82% and 84% for the Distance-Weighted KNN, 99%, 70% and 65% for the ANFIS and 98%, 35% and
95% for the traditional decision tree. With the NLOS detection and exclusion, the proposed GBDT with
multi-feature based method can provide a positioning accuracy improvement of 34.1% compared to
the traditional C/N0 based method.
© 2019 Elsevier B.V. All rights reserved.

1. Introduction signals received together; and ⃝

3 non-line-of-sight (NLOS): where
the user can only receive a reflected signal. Positioning errors
Global navigation satellite systems (GNSS), especially the caused by NLOS and multipath cannot be removed by differential
global positioning system (GPS), are widely used in positioning, techniques [1,2] with the latter potentially resulting in position-
navigation and timing (PNT), increasingly playing an important ing error of around 100 m [3]. Hence, many methods have been
role in every aspect of our lives. The development of smart proposed to mitigate the effects of multipath and NLOS, includ-
bus systems, emerging unmanned aerial vehicles (UAV) and au- ing antenna design, signal processing and measurement-based
tonomous vehicles all require accurate and reliable GPS position- modelling.
ing solutions. GPS signals, however, are easily reflected or blocked Some expensive high-grade antennas could be used to miti-
by buildings, with the consequences of either reduced accuracy gate multipath and NLOS effects [4–6]. Choke-ring antennas can
or no positioning solutions in some urban environments such effectively reduce the multipath effect at low elevation. However,
as canyons. There are three GPS signal reception types: ⃝ 1 line- they are not only expensive but also bulky and hence, inap-
of-sight (LOS): direct signal between the satellite and the user propriate for some location-based applications. Dual-polarization
receiver; ⃝2 multipath: reflected and direct line-of-sight (LOS) antennas could be used for NLOS detection. However, they are
mainly used for geodetic applications such as snow depth detec-
∗ Corresponding author at: College of Civil Aviation, Nanjing University of tion, due to the high cost and bulk of the antennas. Dierendonck
Aeronautics and Astronautics, Nanjing 210016, China. et al. showed how narrowing the spacing between early and late
E-mail address: [email protected] (R. Sun). receiver code correlators helps to mitigate multipath and NLOS

https://doi.org/10.1016/j.asoc.2019.105942
1568-4946/© 2019 Elsevier B.V. All rights reserved.

Please cite this article as: R. Sun, G. Wang, W. Zhang et al., A gradient boosting decision tree based GPS signal reception classification algorithm, Applied Soft Computing
Journal (2019) 105942, https://doi.org/10.1016/j.asoc.2019.105942.
2 R. Sun, G. Wang, W. Zhang et al. / Applied Soft Computing Journal xxx (xxxx) xxx

effects [7]. Subsequently, multipath estimating delay lock loops used in recent years to improve the accuracy of GNSS positioning
(MEDLL) [8], vision correlators and strobe correlators have been and signal reception classification. Phan et al. used elevation
developed [9–13]. These methods, referred to as signal processing and azimuth angles as the key features of support vector ma-
based multipath mitigation, assume that both direct and reflected chines (SVM) to mitigate the multipath effect [37]. Yozevitch
signals are able to reach the receiver, thus a signal processing et al. used C/N0 , elevation and other observations as features
technique can be applied to filter the reflected signal. Unfortu- of decision trees to classify LOS and NLOS [32]. Monsak et al.
nately, these methods do not mitigate the effect of NLOS signal proposed a method using machine learning to detect NLOS signals
reception as it comprises of the reflected signal only. in a collaborative vehicle environment. The results are com-
Measurement-based modelling refers to using satellite ob- pared with several different machine learning algorithms [38].
servations and related information to mitigate NLOS reception Hsu et al. applied SVM to distinguish received GPS signal types,
using four features, including C/N0 , temporal difference of C/N0 ,
and multipath effects. GPS measurements can be augmented by
pseudorange residual and pseudorange rate, extracted from raw
complementary additional sensors (such as inertial measurement
measurements [36]. Quan et al. proposed a convolutional neu-
units) to improve performance in urban canyons. However, this
ral network (CNN) based multipath detection method with the
is highly dependent on the performance of the inertial measure-
sparse auto-encoder (SAE) for feature extraction [39]. Guermaha
ments and their mechanizations [14–21]. Integrating GNSS with
et al. proposed a GNSS signal classifier system based on the
vision sensors can be effective in reducing multipath error and satellite elevation and the difference of C/N0 value, provided by
resulting in the improvement of positioning accuracy. However, right-hand circular polarized (RHCP) and left-hand circular polar-
the performance is affected by weather and environment fea- ized (LHCP) antennas, with the decision tree [40]. Sun et al. used
tures [22]. In addition, the expense associated with such sensors nine variables, including the C/N0 , temporal difference of C/N0 ,
precludes their use in location-based services. HDOP, VDOP, satellite elevation angle and azimuth angle, pseu-
Another relatively low-cost approach to mitigate multipath dorange residual, consistency between delta pseudorange and
errors is to use a 3D city model as an information source. The pseudorange rate and number of visible satellites, derived from
model can be used to classify the type of signal reception and thus the raw GPS measurements together with an algorithm based on
exclude NLOS signals during positioning [23,24]. However, this is an ANFIS to classify LOS, multipath and NLOS measurements from
very likely to decrease the number of available satellites, leading GPS [41]. These studies have shown that machine learning is a po-
both to a weaker geometry and lower redundancy required for tentially effective method for GPS signal reception classification.
the reduction in random errors and integrity monitoring. A more To date, this potential is still to be exploited, because it is difficult
reasonable approach is to use the NLOS signal for positioning to obtain a trade-off between high classification accuracy and low
with the shadow matching method. Shadow matching utilizes 3D computational cost with an increasing number of input features.
city models to predict satellite visibility and then compares this Therefore, in this paper we are aiming to use a number of in-
prediction with the measured satellite visibility to determine the put features with various machine learning algorithms to improve
position [25–27]. In addition, a research team at the University classification accuracy with a high computational efficiency. The
of Tokyo used a 3D city model to simulate signal propagation main known features from GPS raw measurements include C/N0 ,
path and computed simulated pseudorange measurements, be- HDOP, VDOP, satellite elevation angle, azimuth angle, pseudor-
ange residual, pseudorange rate and number of visible satellites.
fore weighting each candidate based on the similarity between
This paper proposes a signal reception classifier based on the gra-
the simulated and actual pseudorange measurements. The esti-
dient boosting decision tree (GBDT) algorithm using 3 variables
mated position was then obtained by the weighted average of the
as features: C/N0 , pseudorange residuals, and satellite elevation
candidate positions [28–30]. The performance of this method de-
angle. The reason for choosing these 3 features among the known
pends on the correct and reliable classification of LOS, multipath
features are that C/N0 , pseudorange residuals, and satellite eleva-
and NLOS signals. tion angle are highly related to the types of signal reception, and
The traditional method for signal classification is to define a their ability to distinguish the signal reception type have been
threshold value of C/N0 with higher C/N0 classified as LOS, while proved by previous studies [31–40]. For the other features, such
those with lower C/N0 classified as NLOS. Although Yozevitch as HDOP, VDOP and GDOP, represent the geometric distribution
et al. have shown that under conditions of no interference, C/N0 of visible satellites in a single epoch without a direct relationship
can serve as an excellent classifier [31,32], in practice LOS signals to signal reception type. The pseudorange rate only represents
can be detected with low C/N0 for various reasons (e.g., the an instantaneous change of the signal and therefore, has a weak
antenna location, momentary blocks, etc.) and NLOS signals can relationship to signal reception type. The contributions of this
be detected with relatively high C/N0 . Hence, additional signal paper are summarized as follows.
features are required for reception type classification. Therefore,
in addition to C/N0 , satellite elevation can be used as a feature for • A robust GBDT based GPS signal reception classification
the classification. In general, the higher the elevation, the more algorithm is proposed by using C/N0 , pseudorange residuals
likely it is for a signal to be LOS, but as the elevation decreases, the and satellite elevation angle as the input features to improve
satellite is more likely to be blocked by buildings and other man- the performance of GPS signal reception classification.
made obstacles, with the corresponding signal being NLOS. Deng • A sensitivity analysis process is designed and used to select
proposed a satellite selection algorithm based on satellite eleva- the most appropriate iteration number of the algorithm
to ensure the classification accuracy, while avoiding over
tion angle and geometric dilution precision (GDOP). The influence
fitting of the GBDT algorithm
of elevation angle on positioning accuracy is analysed [33]. Wang
et al. also use C/N0 and satellite elevation angle to evaluate the
• The raw GPS measurements are collected in various urban
canyon environments to evaluate the performance of the
possibility of LOS [34]. In addition, larger pseudorange residuals
designed algorithm
could point to a higher probability of NLOS or multipath, and
therefore, this can also be used to classify signal reception [35]. 2. Algorithm design
The other features that could be used for classification are pseu-
dorange change rate, horizontal dilution of precision (HDOP) and The framework of the algorithm (Section 2.1) consists of the
vertical dilution of precision (VDOP) [32,36]. three functions of data labelling (Section 2.2), feature selection
Machine learning has the advantages of high speed and ac- (Section 2.3) and the GBDT based signal classification algorithm
curacy in dealing with various types of features and has been (Section 2.4).

Please cite this article as: R. Sun, G. Wang, W. Zhang et al., A gradient boosting decision tree based GPS signal reception classification algorithm, Applied Soft Computing
Journal (2019) 105942, https://doi.org/10.1016/j.asoc.2019.105942.
 R. Sun, G. Wang, W. Zhang et al. / Applied Soft Computing Journal xxx (xxxx) xxx 3

Fig. 1. Classification algorithm functions.

2.1. Algorithm framework

The process of the machine learning based LOS/Multipath/

NLOS classification algorithm, including the offline and online
parts, is presented in Fig. 1. For the offline process, the dataset
containing a large number of LOS, NLOS and Multipath signals is
used for training. The LOS signals are labelled and obtained from
a GPS reference station located in an open area. With the aid of a
3D city model, GPS measurements collected from urban canyons
are classified and labelled as either multipath or NLOS by using a
ray-tracing algorithm. The five candidate machine learning algo-
rithms are used separately to train the labelled training dataset
for the extraction of classification rules. The extracted rules are
then used in the classification of newly collected unlabelled GPS
measurements in the online process.

2.2. Data labelling based on the 3D city model and ray-tracing

A 3D city model is an important information source for the

algorithm. In this paper, the horizontal coordinates of the 3D city Fig. 2. Ray-tracing technique.
model were from the topographic map from the Land Department
of the HK government with a resolution of a 20 cm. The height
of the 3D city model was determined from Google map plus the First calculate an intermediate vector t 0 before calculating the
height of the equipment. position r, as shown in the following equation:
With the building corner coordinates from the 3D city model,
the ray-tracking method was applied to distinguish between the t 0 = ((c − p) · n) /(n · n) (2)
multipath and NLOS measurements in the datasets collected from c can be any point on the plane, here it is taken as one of the
urban canyons. The ray-tracing method uses known satellite, vertices on the plane. After finding the intermediate vector t 0 ,
reflector and receiver geometry to trace the direct and reflected calculate the position r:
paths [42]. Satellite positions are obtained from the broadcast
ephemeris. The position of the reflectors are derived from a 3D r = p + t 0n (3)
building model. The principle of ray-tracing is shown in Fig. 2. (2) Find the mirror image position vector q of the receiver
Suppose that A, B, C and D are four vertices of a building, and antenna position p relative to the reflection surface:
the corresponding position vectors are a, b, c and d respectively.
Since Fig. 2 is a side view of the reflector, only c and d are marked q = p + 2(r − p) (4)
in the picture. The normal vector n of the surface can be obtained (3) Connect the satellite position g to the mirror image q of
by taking the cross product of two non-parallel vectors formed by the antenna and find the intersection point s of the line segment
the vertices of the surface. For example, using vertices A, B and and plane consisting of the reflection surface.
C, n can be calculated by:
s = g + t (q − g ) (5)
n = (b − a) × (c − a) (1)
where t is an intermediate quantity expressed as:
The steps of ray tracing are:
t = ((c − g ) · n) /((q − g ) · n) (6)
(1) Find the position r closest to the receiver antenna position
p on the reflecting surface and calculate the position vector (4) Determine if the intersection is in the reflection surface.
difference r − p between the two points. If the intersection is outside of the surface, prepare two line

Please cite this article as: R. Sun, G. Wang, W. Zhang et al., A gradient boosting decision tree based GPS signal reception classification algorithm, Applied Soft Computing
Journal (2019) 105942, https://doi.org/10.1016/j.asoc.2019.105942.
4 R. Sun, G. Wang, W. Zhang et al. / Applied Soft Computing Journal xxx (xxxx) xxx

segments to connect the point of reflection and the satellite, and decreases, however, there is a more likelihood of signal blockage
the point of reflection and the receiver. If both line segments are by buildings and other obstacles. Elevation angle can, therefore,
not blocked by some other structure, they are considered as a be used as a feature in signal reception classification. The satellite
reflected path. If the intersection is outside the surface, a reflected elevation angle θ can be calculated as (11):
path does not exist. If the receiver only receives a reflected signal
from a satellite, the measurements are labelled as NLOS. If the
θ (i) = − arcsin(u(Di) ) (11)
receiver receives both direct and reflected signals from a satellite,
2.4. Signal classification algorithm based on GBDT
the measurements are labelled as multipath.
GBDT is a supervised learning algorithm [45], also known
2.3. Feature selection
as gradient boost regression tree (GBRT) and multiple additive
regression tree (MART). It combines regression trees using a gra-
Currently, the signal features can be easily obtained from
dient boosting technique and has been widely applied in various
modern GNSS receivers. The features selected in this paper are:
disciplines, such as credit risk assessment [46], transport crash
C/N0 , pseudorange residuals and satellite elevation angle.
prediction [47] and fault prognosis in electronic circuits [48]. It
(1) C/N0 : the signal strength is measured in terms of the C/N0 ,
replaces the difficult function minimization problem by using
which is the ratio of carrier power to noise power per unit of
least-squares function minimization, followed by only a single
bandwidth in decibel-hertz (dB-Hz). Usually, the C/N0 of an NLOS
parameter optimization based on the original criterion. Therefore,
signal is smaller than that of a LOS signal. Therefore, it is the most
this advantage could potentially facilitate the achievement of high
commonly used feature. However, since both low and high C/N0
accuracy GPS signal reception classification [45].
values are possible for NLOS in urban canyons due to the different
In the designed GBDT based algorithm, each sample in the
reflection surface materials, classification simply based on C/N0 training set is represented as xi = (C /N0i , ηi , θi ), where i =
can be unreliable and therefore, necessitates the consideration of 1, 2, 3, . . . , N indicating the sequence number of the sample, and
additional features. N is the number of samples. The labelled training dataset can
(2) Pseudorange residuals, η: the pseudorange ρ is computed be expressed as T = {(x1 , y1 ) , (x2 , y2 ) , (x3 , y3 ) , . . . , (xN , yN )},
as the time ∆T from the signal being transmitted from the satel- where yi ∈ {−1, 0, 1}, is the label of each sample, −1, 0, 1 rep-
lite to the detection of the signal in the receiver multiplied by the resent the NLOS, Multipath and LOS signals, respectively. GBDT
speed of light c plus the clock synchronization error t multiplied minimizes the expected value of loss function L (yi , f (xi )) by
by the speed of light c, which can be expressed in (7): iteratively creating a weak learner ht (xi ; a) that points in the
ρ = ∆T × c + t × c (7) steepest-descent direction, i.e., the negative gradient direction.
The weak learner ht (xi ; a) is a classification tree, the parameters
The satellite position can be resolved from the broadcast a are the splitting variables, split locations and the terminal node
ephemeris, so the positioning solution can be calculated by solv- means of the individual trees. The square loss function (12) is
ing the pseudorange equations using least square estimation in used in this paper:
(8):
1
)−1 L (yi , f (xi )) = (yi − f (xi ))2 (12)
r = GT G GT ρ
(
(8) 2
The input to GBDT is the labelled training dataset T , with M
where r is the receiver state, including the three-dimensional
as the number of iterations. The GBDT based GPS signal reception
position and the receiver clock offset. G denotes the design ma-
(i) (i) (i) classification algorithm flow is as follows:
trix consisting of the unit LOS vector (uN , uE , uD ) between the
1. Initialize a weak learner f0 (x) for the training data:
satellite and receiver in (9):
N
(1) (1) (1)
⎡ ⎤ ∑
uN uE uD −1 f0 (x) = argmin L(yi , γ ) (13)
⎢ (2) (2) (2) ⎥ γ
⎢ uN uE uD −1⎥ i=1
G=⎢
⎢ .. .. .. .. ⎥
⎥ (9)
⎣ . f0 (x) is a regression tree consisting of only one root node. Since
. . . ⎦
(i) (i) (i) L is selected to be the square loss function, f0 (x) becomes:
uN uE uD −1
f0 (x) = y (14)
Once the positioning solution is calculated, the distance be-
tween the receiver and satellite can be obtained. The difference 2. For m = 1 to M:
between this distance and the pseudorange is called the pseudo- 2.1 Compute the negative gradient
range residual, expressed as η, which can be calculated in (10): [
∂ L (yi , f (xi ))
]
ỹi = − (15)
∂ f (xi ) f (x)=fm−1 (x)
η =ρ−G·r (10)
2.2 Replace the label{(yi of the
) training ) dataset )with ỹi(to obtain
Pseudorange residuals are important for LOS/Multipath/NLOS x1 , ỹ1 , x2 , ỹ2 , x3 , ỹ3 , . . . , xN , ỹN ,
( ( )}
a new dataset Tm =
signal reception classification [43]. In theory, the absolute value of and create a new regression tree hm (xi ; am ) by training the new
pseudorange residuals and the probability of NLOS are positively dataset Tm :
related. This phenomenon becomes more obvious when only a N
small portion of the signals are NLOS [44]. Hsu et al. showed ∑
at = argmin (ỹi − hm (xi ; a))2 (16)
that the pseudorange residuals could be used as an indicator to a
i=1
classify signal reception type if the number of measurements is
sufficient [35]. 2.3 Update the strong learner:
(3) Satellite elevation, θ : there is a significant positive cor-
fm (x) = fm−1 (x) + ρ hm (x; am ) (17)
relation between satellite elevation and probability of LOS. In
general, signals from satellites with higher elevation angles are where ρ is the learning rate, usually chosen to be a value between
less likely to be blocked or reflected by buildings. As the elevation 0∼1 to prevent the overfitting.

Please cite this article as: R. Sun, G. Wang, W. Zhang et al., A gradient boosting decision tree based GPS signal reception classification algorithm, Applied Soft Computing
Journal (2019) 105942, https://doi.org/10.1016/j.asoc.2019.105942.
 R. Sun, G. Wang, W. Zhang et al. / Applied Soft Computing Journal xxx (xxxx) xxx 5

Fig. 3. Data collection and signal reception classification process.

3. After the iteration is terminated, output fM (x) as the final

classifier:
M
∑
fM (x) = f0 (x) + ρ hm (x; am ) (18)
m=1

4. fM (x) is used to predict the signal reception type of the newly

collected unlabelled sample x = (C /N0 , η, θ ) from the testing
dataset. The predicted values need to be rounded to the closest
value of 1, 0 or −1.

3. Field test and analysis of results

The proposed GPS signal reception classification algorithm

based on the GBDT is compared with current classification al-
gorithms, including decision tree, distance weighted k-nearest
neighbour (KNN) and the adaptive network-based fuzzy inference
system (ANFIS). Decision tree uses a tree structure learning as
a predictive model to go from observations about an item to Fig. 4. GPS data collection environment in a built-up area.
conclusions about the item’s target value. Distance weighted KNN
is an improved of KNN by heavier weighing of the close neigh-
bours, according to their distances to the query. ANFIS integrates
were both in the urban canyon, it was assumed that only NLOS
neural network (NN) training with fuzzy inference system (FIS),
and multipath signals were present. The labelling of NLOS and
taking linguistic rules from human experts, and adapting itself
multipath was also obtained by using a 3D city model and ray-
by using the input–output data to obtain better training perfor-
tracing. In one case, however, it was difficult to obtain labelled
mance [32,41,49]. The Experiment is presented in Section 3.1 and
data due to the limitations of the 3D city model used with the
the resulting data analysed in Section 3.2.
ray-tracing-technique.
Dataset D0 was created by combining the data collected from
3.1. Experimental process Location A and Location R. D0 was subsequently used to generate
two datasets: training set D1 and testing set D2. In order to
Five datasets were collected from four different locations prevent bias in the training results due to uneven sample distri-
(Fig. 3). Static GPS data were captured for a period of 24 h at bution, an equal number of LOS, NLOS, and Multipath samples
an interval of 30 s at the SatRef HKSC reference station (Lo- were randomly selected from D0 to form the training dataset
cation R), Hong Kong. The data collected from the Location R D1, which contained 24,000 samples. Thus, in D1, the labelling
was labelled as LOS. In addition, static GPS data were captured distribution for each of LOS, NLOS, and multipath was one third.
over 24 h at the same interval in a built-up area in Hung Hom Only 24,000 of the 96,992 samples were selected for training,
(Location A) using a commercial GPS receiver, u-blox NEO-M8T, to reduce the computational load and prevent overfitting. There
as shown in Fig. 4. The dataset from Location A contained pre- were 18 164 multipath samples in D0 and 8000 were used in
dominantly multipath and NLOS measurements. Additional static the training dataset. Therefore, 8000 samples of each type were
GPS data were collected from different urban canyon environ- randomly selected from the remainder of D0 (i.e. excluding D1)
ments (Location B and Location C). Since Locations B and C to form the testing dataset D2 in order to evenly distribute the

Please cite this article as: R. Sun, G. Wang, W. Zhang et al., A gradient boosting decision tree based GPS signal reception classification algorithm, Applied Soft Computing
Journal (2019) 105942, https://doi.org/10.1016/j.asoc.2019.105942.
6 R. Sun, G. Wang, W. Zhang et al. / Applied Soft Computing Journal xxx (xxxx) xxx

Table 1
Summary of the datasets.
Dataset D0 D1 D2 D3 D4
Total samples 96 992 24 000 24 000 11 615 25 039
LOS (labelled as 1) 25 987 8 000 8 000 0 0
Multipath (labelled as 0) 18 164 8 000 8 000 3 114 8 831
NLOS (labelled as −1) 52 841 8 000 8 000 8 501 16 208

three types of samples in the dataset. Thus, the labelling dis-

tribution of D2 for each of LOS, NLOS, and multipath was also
one third, but with the labels removed. Each measurement in D1
and D2 contained the associated three features of: C/N0 , pseu-
dorange residual and satellite elevation angle. Although some of
the features could be correlated over time, in this case, the time
dependency of the data was not considered when applying GBDT
algorithm.
We used training dataset D1 with each of the four machine
Fig. 5. Relationships between the features (i.e. C/N0 , satellite elevation angles
learning algorithms to determine classification rules. The clas-
and pseudorange residuals) and the corresponding labelled signal reception
sification rules extracted by the machine learning algorithms types in D0.
were then used to classify the samples in D2. These were then
compared to the classification results with the reference label for
each sample (i.e. the previously removed labels), to calculate the
accuracy of the algorithm. To verify the validity of the extracted
rules, two more testing datasets collected from other locations
were used to feed the rules. Testing dataset D3 was collected
from Location B, close to location A in the urban canyon, while
the other testing dataset, D4, was from Location C, about three
blocks away from location A in the urban canyon. A summary of
the datasets is presented in Table 1.

3.2. Results and analysis

The GBDT algorithm was used to train dataset D1 to deter-

mine classification rules, subsequently used to classify the testing
datasets to determine classification accuracy. The results were
then compared to those from the three other machine learning
algorithms. The candidate machine learning algorithms could be
used with either single or multiple signal features. To determine
the benefit of the additional features, a comparison was made
between classification using C/N0 only and the multiples features Fig. 6. Relationship between the features (i.e. C/N0 and satellite elevation angles)
of C/N0 , pseudorange residuals and satellite elevation angle. and the corresponding labelled signal reception types in D0.
The confusion matrix of the LOS, multipath and NLOS (1, 0
and −1) classification results for the different algorithms using
single feature-based classification (i.e. C/N0 ), by using the testing C/N0 areas. The large overlaps makes it difficult to distinguish
dataset D2 are compared in Table 2. The accuracy in the table NLOS from multipath signals by using C/N0 only.
represents the ratio (in percentage) of the number of samples Fig. 6 shows the relationship between the features (i.e. C/N0
correctly classified to the total number of samples in the dataset.
and satellite elevation angles) and the corresponding labelled
The accuracy of each category refers to the ratio (in percentage) of
signal reception types in dataset D0. It can be seen that lower
the number of samples correctly classified to the total number of
elevation angles are dominated by NLOS signals, with most of
samples known as being in this category. For example, the NLOS
their C/N0 values less than 45dB-Hz. Multipath accounts for the
detection accuracy is calculated as the ratio (in percentage) of the
majority of the samples with elevation angles higher than 25
number of samples correctly classified as NLOS to the number
degrees.
of total known NLOS samples. For single feature (C/N0 ) based
classification, the classification accuracies of the four algorithms Fig. 7 shows the relationship between features (i.e. C/N0 and
for multipath and NLOS, are consistently below 80% (ranging from pseudorange residuals) and the corresponding labelled signal re-
29% to 79.6%). The classification accuracy of LOS signals is higher ception types in dataset D0. The pseudorange residuals of LOS
ranging from 95.1% to 99.9%. signals range from −2 m to 2 m with a mean value close to zero.
As discussed earlier, additional features could be used to im- The pseudorange residuals for multipath signals largely range
prove accuracy the traditional single feature, C/N0 based classi- from −100 m to 100 m, while for NLOS signals the residuals are
fication. The scatter diagram in Fig. 5 shows the relationships always over 100 m.
between the input features (i.e. C/N0 , satellite elevation angle and From the analysis of the features and their corresponding
pseudorange residuals) and their corresponding labelled signal labelled signal reception types, consideration of the additional
reception types in the dataset D0. The green points denoting the features (satellite elevation angle and pseudorange residuals) has
LOS signals are concentrated in high C/N0 areas, while the NLOS the potential to improve the accuracy of the GPS signal reception
and multipath samples are mainly distributed in medium and low classification.

Please cite this article as: R. Sun, G. Wang, W. Zhang et al., A gradient boosting decision tree based GPS signal reception classification algorithm, Applied Soft Computing
Journal (2019) 105942, https://doi.org/10.1016/j.asoc.2019.105942.
 R. Sun, G. Wang, W. Zhang et al. / Applied Soft Computing Journal xxx (xxxx) xxx 7

Table 2
Confusion matrix of LOS (Noted as 1), Multipath (0) and NLOS (−1) classification results using
different algorithms based on a single feature (C/N0 ) for Testing Dataset D2.
Algorithms GBDT Decision tree
Label ID −1 0 1 −1 0 1
−1 6365 1635 0 3729 4018 253
No of samples 0 4197 3803 0 1836 4611 1553
1 0 391 7609 0 182 7818
Accuracy (%) 74.1 67.3
Category accuracy (%) 79.7 47.5 95.1 46.6 57.6 97.7
Algorithms Distance-weighted KNN ANFIS
Label ID −1 0 1 −1 0 1
−1 15 7735 250 6054 1691 255
No. of samples 0 0 6170 1830 3934 2329 1737
1 0 22 7978 0 365 7635
Accuracy (%) 59 66.7
Category accuracy (%) 69.4 29 99.9 75.7 29.1 95.4

Fig. 8. Relationship between the number of iterations and the GBDT

Fig. 7. Relationships between features (i.e. C/N0 and pseudorange residuals) and classification accuracy.
the corresponding labelled signal reception types in dataset D0.

affects the adaptively of classification accuracy for the additional

The selection of an appropriate number of iterations, i.e. the datasets. Therefore, in order to keep a high classification accu-
number of regression trees, is very important for the GBDT al- racy without overfitting, the number of iterations is set to 100
gorithm. If the number of iterations is too small, the trained in this paper. In addition, the classification accuracies of GBDT
classification rules cannot fully determine the relationships be- for datasets D3 and D4 are always lower than that of D1 and
tween input features and signal reception types. On the other D2, reflecting the sensitivity of classification rules to different
hand, if there are too many iterations, the results will be prone locations.
to overfitting and the computational load is increased. Therefore, A comparison of the classification results for different algo-
in this paper, before comparing GBDT with other algorithms, a rithms using C/N0 , pseudorange residuals and satellite elevation
sensitivity analysis of the relationship between the number of angle based on the testing dataset D2, is presented in Table 3.
iterations and classification accuracy of each dataset is carried The classification accuracy and training time of these algorithms
out. The classification rules determined from different numbers are shown in Figs. 9 and 10, respectively. By comparing the
of training iterations with training dataset D1 are used to classify results from Tables 2 and 3, it is obvious that the classification
the signal reception types of the dataset D1 for internal valida- accuracies of multiple feature-based algorithms are higher than
tion (self-consistency check) and testing sets D2, D3, and D4 for that of single feature (C/N0 ) based algorithms. In addition, Table 3
external validation (testing). The results are presented in Fig. 8. shows that the overall classification accuracy of the GBDT algo-
Overall the classification accuracy increases with the number rithm with the three features 89%, better than that with decision
of iterations. The classification accuracy of D2 has been increased tree and ANFIS, and marginally higher than that with distance-
greatly as it is from the same data source as the dataset D1. This is weighted KNN. Although the NLOS classification accuracy of the
the reason why the classification rules based on the large number GBDT algorithm is lower than that of traditional decision tree, the
of iterations obtained from D1 are applicable to D2. For the testing latter has a higher number of missed detections, with multipath
datasets D3 and D4 from different locations compared to D1, signals identified as NLOS. In the iterative process, GBDT focuses
the classification accuracy increases initially with the increasing on the samples with large training residuals (i.e. samples that are
iteration numbers but then starts to decline when the iteration difficult to classify). The final learner is the fusion of multiple re-
number reaches a certain value (i.e. 100) due to overfitting of gression trees created in each iteration with overfitting mitigated
the algorithm. Although the more the number of iterations the to some extent by setting the weight of each regression tree based
better the fitting of the extracted rules for the training dataset, it on the learning rate. For the decision tree, the training dataset can

Please cite this article as: R. Sun, G. Wang, W. Zhang et al., A gradient boosting decision tree based GPS signal reception classification algorithm, Applied Soft Computing
Journal (2019) 105942, https://doi.org/10.1016/j.asoc.2019.105942.
8 R. Sun, G. Wang, W. Zhang et al. / Applied Soft Computing Journal xxx (xxxx) xxx

the validity of the extracted classification rules. The classification

accuracies for the testing datasets D3 and D4 are compared in
Tables 4 and 5.
The classification accuracy performance based on the testing
dataset D3, from a similar urban environment to the training
data, is illustrated in Table 4. The GBDT based algorithm has
an overall classification accuracy of 77%, which is higher than
that of distance-weighted KNN (i.e. 68%) and ANFIS (i.e. 71.5%).
Although the decision tree-based algorithm has a relative higher
overall classification accuracy (i.e. 86.1%), the classification accu-
racy for the multipath is low (i.e. 54%). More multipath signals
are misclassified as NLOS signals by using the decision tree.
The performance based on the testing dataset D4, from a
different urban environment to the training data, is illustrated in
Table 5. The overall classification accuracy for all the candidate
algorithms, including the GBDT, are around 55%–60%, reflecting
the data sensitivity of the machine learning algorithms.
We have further evaluated the static positioning results based
on the elimination of the NLOS signals detected from the pro-
posed GBDT with multi-feature-based method and single C/N0
based method. Here, the threshold of single C/N0 is also deter-
mined by GBDT, which has a higher classification results com-
pared to the other algorithms in Table 2. The comparisons of the
static positioning accuracy Root Mean Square Error (RMSE) of the
dataset D2, D3 and D4 from Location A, B and C, respectively, are
Fig. 9. Classification accuracy for different algorithms using multiple features
(C/N0 , Pseudorange Residuals and Satellite Elevation Angle) of the testing dataset
shown in Table 6 and the comparisons of positioning results for
D2. the three locations are shown in Figs. 11, 12 and 13, respectively.
For the Location A, where the urban training dataset is col-
lected, the 2D and 3D RMSE of the positioning accuracy for the
proposed method are 31.23 m and 59.82 m, 30.1% and 34.1%
higher than that of the single feature C/N0 based NLOS elimina-
tion method. As shown in Fig. 11, although positioning results for
both algorithms failed to cover ground truth due to the low mea-
surement source, compared with the single feature C/N0 based
NLOS elimination method, the positioning results of the proposed
method are more concentrated and closer to the ground truth.
For the Location B, which is not far away from location A, the
2D and 3D RMSE of the positioning results from the proposed
method are 44.06 m and 64.77 m, 19.4% and 19.8% higher than
that of the single feature C/N0 based NLOS elimination method.
The positioning results of the proposed method are closer to
the ground truth compared to the single C/N0 based method as
depicted in Fig. 12. For the Location C, from a different urban
canyon environment, the positioning accuracy improvement from
the GBDT with multi-feature based method is limited, with an
improvement of 2.2% for the 2D and an deterioration of 5.9% for
the 3D positioning results, compared to the single C/N0 based
method due to the similar low performance of the signal recep-
tion classification (i.e. about 55% for multi-feature and 57% for
Fig. 10. Training time for different algorithms using multiple features (C/N0 ,
Pseudorange Residuals and Satellite Elevation Angle) of the testing dataset D2.
single feature). It is indicated from the analysis results that for the
Location A and Location B, the proposed method can effectively
classify the signal reception types, and therefore improve the
positioning results after removal of the NLOS signal detected.
be better fitted by increasing the number of leaf nodes. However, While for the Location C, the positioning accuracy could not be
as the complexity of the tree increases, overfitting is more likely improved due to the poor classification performance from the
to occur and leads to more misclassification. Therefore, the GBDT proposed algorithm. It is validated that the higher accuracy of the
composed of multiple regression trees is better than a single signal reception classification could result in more improvement
complex decision tree. Although the classification accuracy of of the final GPS positioning results.
weighted KNN is close to GBDT and with a lower training time To be noted that the positioning based on the NLOS exclusion
consumption, the classification rules from the KNN-algorithm are could result in a 20% to 35% improvement based on proposed
less adaptable to other datasets. The classification accuracy of AN- algorithm for the similar urban environments. It is potential to
FIS is over 80% but with a high misclassification rate between the further improve the positioning results with the correction of
multipath and NLOS measurements. The ANFIS based algorithm the detected NLOS signals used in the positioning process. The
also has the longest training time among all of the candidate labelling error caused by the inaccuracy borders of the 3D map is
algorithms. also a reason to affect the final positioning accuracy. Therefore,
Dataset D3 and dataset D4 collected from location B and loca- the positioning accuracy could be further improved with the
tion C in different urban canyon areas are further used to verify improved high-definition 3D map.

Please cite this article as: R. Sun, G. Wang, W. Zhang et al., A gradient boosting decision tree based GPS signal reception classification algorithm, Applied Soft Computing
Journal (2019) 105942, https://doi.org/10.1016/j.asoc.2019.105942.
 R. Sun, G. Wang, W. Zhang et al. / Applied Soft Computing Journal xxx (xxxx) xxx 9

Table 3
Confusion matrix of LOS (1), Multipath (0) and NLOS (−1) classification results using different
algorithms based on multiple features (C/N0 , Pseudorange Residuals and Satellite Elevation Angle)
for Testing Dataset D2.
Algorithms GBDT Decision tree
Label ID −1 0 1 −1 0 1
−1 6858 1134 8 7583 395 22
No of samples 0 1322 6522 156 5006 2795 199
1 0 14 7986 133 5 7862
Accuracy (%) 89 76
Training Time (s) 47.6 3.1
Category accuracy (%) 85.7 81.5 99.8 94.8 34.9 98.3
Algorithms Distance-weighted KNN ANFIS
Label ID −1 0 1 −1 0 1
−1 6521 1472 7 6256 1716 28
No. of samples 0 1266 6724 10 2237 5634 129
1 0 1 7999 0 47 7953
Accuracy (%) 88.5 82.7
Training time (s) 1.2 105
Category accuracy (%) 81.5 84.1 100 64.9 70.4 99.4

Table 4
Confusion matrix of LOS (1), Multipath (0) and NLOS (−1) classification results using different
algorithms based on multiple features (C/N0 , Pseudorange Residuals and Satellite Elevation Angle)
for Testing Dataset D3.
Algorithms GBDT Decision tree
Label ID −1 0 1 −1 0 1
−1 6239 2262 0 8320 168 13
No of samples 0 299 2726 89 1321 1681 112
1 0 0 0 0 0 0
Accuracy (%) 77.2 86.1
Category accuracy (%) 73.4 87.5 97.9 54
Algorithms Distance-weighted KNN ANFIS
Label ID −1 0 1 −1 0 1
−1 6217 2284 0 7047 1436 18
No. of samples 0 1024 1687 403 868 1252 994
1 0 0 0 0 0 0
Accuracy (%) 68 71.5
Category accuracy (%) 73.1 54.2 82.9 40.2

Table 5
Confusion matrix of LOS (1), Multipath (0) and NLOS (−1) classification results using different
algorithms based on multiple features (C/N0 , Pseudorange Residuals and Satellite Elevation Angle)
for Testing Dataset D4.
Algorithms GBDT Decision tree
Label ID −1 0 1 −1 0 1
−1 8830 7378 0 14 485 1723 22
No of samples 0 3694 5025 112 7 782 851 198
1 0 0 0 0 0 0
Accuracy (%) 55.3 61.3
Category accuracy (%) 54.6 56.6 89.4 9.6
Algorithms Distance-weighted KNN ANFIS
Label ID −1 0 1 −1 0 1
−1 9217 6991 0 9385 6663 160
No. of samples 0 2970 5775 86 3141 5374 316
1 0 0 0 0 0 0
Accuracy (%) 60 59
Category accuracy (%) 56.9 65.4 57.9 60.9

In summary, the classification rules extracted from the GBDT 4. Conclusions and future work
algorithm are applicable to environments with largely similar
spatial and material characteristics (i.e. testing dataset D2 and D3) This paper has presented a GBDT based algorithm, using C/N0 ,
pseudorange residual and elevation angle as the features, to clas-
but with low adaptively to datasets with different characteristics
sify GPS signal reception as LOS, multipath and NLOS, mean-
(i.e. testing dataset D4). Further work is exploring further the while, the static positioning solutions are also calculated with the
issue of adaptively and the development of real-time on-line detected NLOS eliminations. The signal reception classification
training algorithms. results of the testing dataset D2 (from Location A), from the

Please cite this article as: R. Sun, G. Wang, W. Zhang et al., A gradient boosting decision tree based GPS signal reception classification algorithm, Applied Soft Computing
Journal (2019) 105942, https://doi.org/10.1016/j.asoc.2019.105942.
10 R. Sun, G. Wang, W. Zhang et al. / Applied Soft Computing Journal xxx (xxxx) xxx

Table 6
Comparisons of RMSE for Location A, B and C.
RMSE (m) E N U 3D 2D
C/N0 Based NLOS 40.92 17.9 79.01 90.76 44.67
Location A Elimination
GBDT with 26.19 17.02 51.02 59.82 31.23
Multi-Feature-Based NLOS
Elimination
Improvement (%) 36.0 4.9 35.4 34.1 30.1
C/N0 Based NLOS 20.13 45.41 63.72 80.80 49.67
Location B Elimination
GBDT with 18.35 35.61 50.89 64.77 40.06
Multi-Feature-Based NLOS
Elimination
Improvement (%) 8.8 21.6 20.1 19.8 19.4
C/N0 Based NLOS 25.4 29.5 127.67 133.37 38.59
Location C Elimination
GBDT with 25.07 32.27 123.83 130.39 40.86
Multi-Feature-Based NLOS
Elimination
Improvement (%) 1.3 −9.4 3.0 2.2 −5.9

Fig. 11. Positioning results based on the elimination of the NLOS signals detected from the proposed GBDT with multi-feature-based method and single C/N0 based
method in Location A.

Fig. 12. Positioning results based on the elimination of the NLOS signals detected from the proposed GBDT with multi-feature-based method and single C/N0 based
method in Location B.

Please cite this article as: R. Sun, G. Wang, W. Zhang et al., A gradient boosting decision tree based GPS signal reception classification algorithm, Applied Soft Computing
Journal (2019) 105942, https://doi.org/10.1016/j.asoc.2019.105942.
 R. Sun, G. Wang, W. Zhang et al. / Applied Soft Computing Journal xxx (xxxx) xxx 11

Fig. 13. Positioning results based on the elimination of the NLOS signals detected from the proposed GBDT with multi-feature-based method and single C/N0 based
method in Location C.

same environment as the urban training dataset, have shown positioning accuracy to some extent. However, simple elimina-
that the overall accuracy of the multi-feature-based classification tion cannot meet the requirements of high-precision positioning.
algorithm (i.e. 89% for the static data) is much higher than that of The proposed method is a complement to the existing positioning
the single-feature-based C/N0 classification algorithm (i.e.74.1%). methods based on 3D city model such as shadow matching. In fu-
Furthermore, for the correct classification of signals in the cat- ture studies, we will combine the proposed method with 3D map
egories of NLOS and multipath, the local accuracies are 85.7% to achieve better static positioning accuracy and therefore, could
and 81.5% respectively, better than the decision tree, distance- be used for the civil engineering applications such as building
weighted KNN and ANFIS algorithms. For the testing dataset D3 maintenance, and structural integrity/deformation monitoring in
(from Location B), from the similar environment as the training the urban canyons. For dynamic positioning, research is ongoing
dataset, the algorithm achieves an overall classification accuracy based on a framework of grid of reference points from which data
of 77.2%. In particular, the detection accuracies are 73.4% and will be captured for training. Users will then automatically obtain
87.5% for the NLOS and multipath respectively, better than the the classification rules of nearby reference points for accurate
decision tree, distance-weighted KNN and ANFIS algorithms. The satellite signal reception type classification, to improve position-
computation time for the GBDT is higher than the decision tree ing. In addition, we will develop online data training for use with
and distance-weighted KNN, however, this could be resolved by the GBDT algorithm for real-time applications such as ground
higher computing processing power. Overall, considering compu- vehicles, pedestrians and unmanned aerial vehicles (UAVs).
tation time and classification accuracy, it is shown that GBDT is
the best of the algorithms investigated, for GPS signal reception Declaration of competing interest
classification. It should be noted that, for some datasets with
different characteristics from the training dataset, such as D4 No author associated with this paper has disclosed any po-
(from Location B), the classification performance is degraded due tential or pertinent conflicts which may be perceived to have
to the inapplicability of the rules extracted from the training impending conflict with this work. For full disclosure statements
environment. refer to https://doi.org/10.1016/j.asoc.2019.105942.
Based on the GBDT with multi-feature-based classification re-
sults, the application for the static positioning results are further Acknowledgements
analysed with the detected NLOS elimination. For the Location
A, a positioning accuracy improvement of 34.1% (3D RMSE) has This work is supported by the National Natural Science Foun-
been achieved compared to the single C/N0 based method. For dation of China (Grant No. 41704022, No. 41974033), Natu-
the Location B, the proposed method could also provide an im- ral Science Foundation of Jiangsu Province, China (Grant No.
provement of the positioning accuracy with 19.8% (3D RMSE), BK20170780), China Postdoctoral Science Foundation funded
lower than that for the Location A. While for the Location C, the Project (Grant No. 2017M623360) and Foundation of Graduate
proposed method is unable to improve the positioning accuracy Innovation Center in NUAA, China (Grant No. KFJJ20180719) and
due to the different spatial and material characteristics. Therefore, Specialized Research Fund for Shandong Provincial Key Labora-
environmental sensitivity is a key issue in the application of tory, China (Grant No. KLWH201813).
classification algorithms. This could be addressed by developing
spatio-temporally dynamic algorithms, consideration of more sig- References
nal related features and training data from a number of different
locations in future work. [1] P. Misra, P. Enge, GPS Measurements and error sources, in: Global Po-
For static positioning, the proposed method can be used to sitioning System: Signals, Measurements and Performance, second ed.,
Ganga-Jamuna Press, Lincoln, MA, USA, 2011, pp. 174–177.
detect the NLOS and multipath signals, which could then be [2] P.W. Ward, J.W. Betz, C.J. Hegarty, Interference, multipath, and scintillation,
used in data pre-processing. The experimental results show that in: Understanding GPS: Principles and Applications, second ed., Artech
removing NLOS based on the proposed method can improve the House, Norwood, MA, USA, 2006, pp. 279–292.

Please cite this article as: R. Sun, G. Wang, W. Zhang et al., A gradient boosting decision tree based GPS signal reception classification algorithm, Applied Soft Computing
Journal (2019) 105942, https://doi.org/10.1016/j.asoc.2019.105942.
12 R. Sun, G. Wang, W. Zhang et al. / Applied Soft Computing Journal xxx (xxxx) xxx

[3] G. MacGougan, G. Lachapelle, R. Klukas, K. Siu, L. Garin, J. Shewfelt, G. Cox, [27] L. Wang, P.D. Groves, M.K. Ziebart, GNSS Shadow matching: improving
Performance analysis of a stand-alone high-sensitivity receiver, GPS Solut. urban positioning accuracy using a 3D city model with optimized visibility
6 (3) (2002) 179–195, http://dx.doi.org/10.1007/s10291-002-0029-z. scoring scheme, Navigation 60 (3) (2013) 195–207, http://dx.doi.org/10.
[4] J.M. Tranquilla, J.P. Carr, H.M. Al-Rizzo, Analysis of a choke ring ground- 1002/navi.38.
plane for multipath control in global positioning system (GPS) applications, [28] S. Miura, L.T. Hsu, F. Chen, S. Kamijo, GPS Error correction with pseu-
IEEE Trans. Antennas Propag. 42 (7) (1994) 905–911, http://dx.doi.org/10. dorange evaluation using three-dimensional maps, IEEE Trans. Intell.
1109/8.299591. Transp. Syst. 16 (6) (2015) 3104–3115, http://dx.doi.org/10.1109/TITS.2015.
[5] Z. Jiang, P.D. Groves, NLOS GPS Signal detection using a dual-polarisation 2432122.
antenna, GPS Solut. 18 (1) (2014) 15–26, http://dx.doi.org/10.1007/s10291- [29] L.T. Hsu, Y. Gu, S. Kamijo, 3D Building model-based pedestrian positioning
012-0305-5. method using GPS/GLONASS/QZSS and its reliability calculation, GPS Solut.
[6] H. Liang, T. Walter, P. Enge, and, G.X. Gao, GNSS Multipath and jamming 20 (3) (2016) 413–428, http://dx.doi.org/10.1007/s10291-015-0451-7.
mitigation using high-mask-angle antennas and multiple constellations, [30] Y. Gu, L.T. Hsu, S. Kamijo, GNSS/Onboard inertial sensor integration with
IEEE Trans. Intell. Transp. Syst. 16 (2) (2015) 741–750, http://dx.doi.org/ the aid of 3-d building map for lane-level vehicle self-localization in
10.1109/tits.2014.2342200.
urban canyon, IEEE Trans. Veh. Technol. 65 (6) (2016) 4274–4287, http:
[7] V.A. Dierendonck, P. Fenton, T. Ford, Theory and performance of narrow
//dx.doi.org/10.1109/TVT.2015.2497001.
correlator spacing in a GPS receiver, Navigation 39 (3) (1992) 265–283,
[31] R. Yozevitch, B.B. Moshe, H. Levy, Breaking the 1meter accuracy bound
http://dx.doi.org/10.1002/j.2161-4296.1992.tb02276.x.
in commercial GNSS devices, in: 27th Convention of the Electrical and
[8] R.D.J.V. Nee, J. Siereveld, P.C. Fenton, B.R. Townsend, The multipath
Electronics Engineers in Israel (IEEEI), Eilat, Israel, 2012, pp. 1-5.
estimating delay lock loop: approaching theoretical accuracy limits, in:
[32] R. Yozevitch, B.B. Moshe, A. Weissman, A robust GNSS LOS/NLOS signal
IEEE/ION Position, Location and Navigation Symposium, Las Vegas, NV,
classifier, Navigation 63 (4pp) (2016) 429–442.
USA, 1994, pp. 246-251, http://dx.doi.org/10.1109/ISSSTA.1992.665623.
[33] G. Deng, GPS Satellite selection algorithm based on satellite elevation angle
[9] L. Garin, J.M. Rousseau, Enhanced strobe correlator multipath rejection for
code & carrier, in: ION GPS, Kansas, MO, USA, 1997, pp. 559-568. and GDOP, Digit. Commun. 37 (2) (2010) 47–50, (in Chinese).
[10] B.R. Townsend, P.C. Fenton, A practical approach to the reduction of [34] L. Wang, P.D. Groves, M. Ziebart, Smartphone shadow matching for better
pseudorange multipath errors in a Ll GPS receiver, in: International cross-street GNSS positioning in urban environments, J. Navig. 68 (3)
Technical Meeting of the Satellite Division of the Institute of Navigation, (2015) 411–433, http://dx.doi.org/10.1017/S0373463314000836.
Salt Lake City, UT, USA, 1994, pp. 143-148. [35] L.T. Hsu, H. Tokura, N. Kubo, Y. Gu, S. Kamijo, Multiple faulty GNSS mea-
[11] M.S. Braasch, Performance comparison of multipath mitigating receiver surement exclusion based on consistency check in urban canyons, Sensors
architectures, in: IEEE Aerospace Conference, Big Sky, MT, USA, 2001, pp. 17 (6) (2017) 1909–1917, http://dx.doi.org/10.1109/JSEN.2017.2654359.
3/1309-3/1315. [36] L.T. Hsu, GNSS multipath detection using a machine learning approach, in:
[12] B.R. Townsend, P.C. Fenton, K.J.V. Dierendonck, R.D.J.V. Nee, Performance International conference on intelligent transportation systems, Yokohama,
evaluation of the multipath estimating delay lock loop, Navigation 42 (3) Japan, 2017, pp. 1-6. http://dx.doi.org/10.1109/ITSC.2017.8317700.
(1995) 502–514, http://dx.doi.org/10.1002/j.2161-4296.1995.tb01903.x. [37] Q. Phan, S. Tan, I.V. Mcloughlin, L.D. Vu, A unified framework for GPS
[13] P.C. Fenton, J. Jones, The theory and performance of NovAtel Inc.’s vision code and carrier-phase multipath mitigation using support vector regres-
correlator, in: ION GNSS, Long Beach, CA, USA, 2005, pp. 2178-2186. sion, Adv. Artif. Neural Syst. (2013) 1–14, http://dx.doi.org/10.1155/2013/
[14] F. Wang, J. Chen, S. Gao, K. Tang, X. Meng, Development and sea trial of 240564.
real-time offshore pipeline installation monitoring system, Ocean Eng. 146 [38] M. Socharoentum, H.A. Karimi, Y. Deng, A machine learning approach to
(2017) 468–476, http://dx.doi.org/10.1016/j.oceaneng.2017.09.016. detect non-line-of-sight GNSS signals in Nav2Nav, in: presented at the 23rd
[15] M.G. Petovello, C. O’Driscoll, G. Lachapelle, Weak signal carrier tracking of ITS World Congress, Melbourne, Australia, 10–14 October, 2016.
weak using coherent integration with an ultra-tight GNSS/IMU receiver, [39] Y. Quan, L. Lau, G.W. Roberts, X. Meng, C. Zhang, Convolutional neural
in: presented at the European Navigation Conference, Toulouse, Lauragais, network based multipath detection method for static and kinematic GPS
France, Apr. 23-25, 2008. high precision positioning, Remote Sens. 10 (12) (2018) 2052, http://dx.
[16] M.G. Petovello, G. Lachapelle, Comparison of vector-based software re- doi.org/10.3390/rs10122052.
ceiver implementations with application to ultra-tight GPS/INS integration, [40] B. Guermah, H.E. Ghazi, T. Sadiki, H. Guermah, A robust GNSS
in: ION GNSS, Fort Worth, Texas, 2006, pp. 2977-2989. LOS/multipath signal classifier based on the fusion of information and
[17] A. Soloviev, C. Toth, D. Grejner-Brzezinska, Performance of deeply inte- machine learning for intelligent transportation systems, in: 2018 IEEE
grated GPS/ins in dense forestry areas, J. Appl. Geodesy 6 (1) (2012) 3–13, International Conference on Technology Management, Operations and
http://dx.doi.org/10.1515/jag-2011-0005. Decisions (ICTMOD), Marrakech, Morocco, 2018, pp. 94-100. http://dx.doi.
[18] A. Soloviev, F.V. Graas, Use of deeply integrated GPS/INS architecture org/10.1109/ITMC.2018.8691272.
and laser scanners for the identification of multipath reflections in urban
[41] R. Sun, L.T. Hsu, D. Xue, G. Zhang, W.Y. Ochieng, GPS Signal reception
environments, IEEE J. Sel. Top. Sign. Proces. 3 (5) (2009) 786–797, http:
classification using adaptive neuro-Fuzzy inference system, J. Navig. (2018)
//dx.doi.org/10.1109/JSTSP.2009.2027796.
1–17, http://dx.doi.org/10.1017/S0373463318000899.
[19] V. Milanes, J.E. Naranjo, C. Gonzalez, J.A. Ruiz, T. Pedro, Autonomous vehicle
[42] L. Lau, P.A. Cross, Development and testing of a new ray-tracing approach
based in cooperative GPS and inertial systems, Robotica 26 (5) (2008)
to GNSS carrier-phase multipath modelling, J. Geod. 81 (11) (2007)
627–633, http://dx.doi.org/10.1017/S0263574708004232.
713–732, http://dx.doi.org/10.1007/s00190-007-0139-z.
[20] S. Zhao, Y. Chen, J.A. Farrell, High-precision vehicle navigation in urban
[43] P.D. Groves, Z. Jiang, L. Wang, M.K. Ziebart, Intelligent urban positioning
environments using an mem’s IMU and single-frequency GPS receiver, IEEE
using multi-constellation GNSS with 3D mapping and NLOS signal detec-
Trans. Intell. Transp. Syst. 17 (10) (2016) 2854–2867, http://dx.doi.org/10.
1109/TITS.2016.2529000. tion, in: International Technical Meeting of the Satellite Division of the
[21] K.W. Chiang, Y.W. Huang, An intelligent navigator for seamless INS/GPS Institute of Navigation, Nashville, TN, USA, Sep. 2012, pp. 458-472.
integrated land vehicle navigation applications, Appl. Soft Comput. 8 (1) [44] P.D. Groves, Z. Jiang, Height aiding, C/N0 weighting and consistency
(2008) 722–733, http://dx.doi.org/10.1016/j.asoc.2007.05.010. checking for GNSS NLOS and multipath mitigation in urban areas, J. Navig.
[22] J.I. Meguro, T. Murata, J.I. Takiguchi, Y. Amano, T. Hashizume, GPS Multi- 66 (5) (2013) 653–669, http://dx.doi.org/10.1017/s0373463313000350.
path mitigation for urban area using omnidirectional infrared camera, IEEE [45] J.H. Friedman, Greedy function approximation: a gradient boosting ma-
Trans. Intell. Transp. Syst. 10 (1) (2009) http://dx.doi.org/10.1109/tits.2008. chine, Ann. Statist. 29 (5) (2011) 1189–1232, http://dx.doi.org/10.2307/
2011688. 2699986.
[23] S. Peyraud, D. Bétaille, S. Renault, M. Ortiz, F. Mougel, D. Meizel, F. [46] L. Zhou, GBDT-SVM Credit risk assessment model and empirical analysis of
Peyret, About non-line-of-sight satellite detection and exclusion in a 3D peer-to-peer borrowers under consideration of audit information, Open J.
map-aided localization algorithm, Sensors 13 (1) (2013) 829–847, http: Bus. Manag. 6 (2018) 362–372, http://dx.doi.org/10.4236/ojbm.2018.62026.
//dx.doi.org/10.3390/s130100829. [47] H. Park, A. Haghani, S. Samuel, A.K. Michael, Real-time prediction and
[24] F. Peyret, D. Betaille, M. Ortiz, S. Miquel, L. Fontenay, How to improve avoidance of secondary crashes under unexpected traffic congestion, Accid.
GNSS positioning quality of service for demanding ITS in city environments Anal. Prev. 112 (2018) 39–49, http://dx.doi.org/10.1016/j.aap.2017.11.025.
by using 3D digital maps, in: presented at the 19th ITS World Congress, [48] L. Wang, D. Zhou, H. Zhang, W. Zhang, J. Chen, Application of relative
Vienna, Austria, Oct. 22-26, 2012. entropy and gradient boosting decision tree to fault prognosis in elec-
[25] P.D. Groves, Shadow matching: a new GNSS positioning technique for tronic circuits, Symmetry 10 (10) (2018) 495, http://dx.doi.org/10.3390/
urban canyons, J. Navig. 63 (3) (2011) 417–430, http://dx.doi.org/10.1017/ sym1010049.
S0373463311000087. [49] J.R. Jang, ANFIS: adaptive-network-based fuzzy inference system, Syst. Man
[26] L. Wang, P. Groves, M. Ziebart, Urban positioning on a smartphone: real- Cybern. 23 (3) (1993) 665–685, http://dx.doi.org/10.1109/21.256541.
time shadow matching using GNSS and 3D city models, in: ION GNSS,
Nashville, Tennessee, USA, 2013, pp. 1606-1619.

Please cite this article as: R. Sun, G. Wang, W. Zhang et al., A gradient boosting decision tree based GPS signal reception classification algorithm, Applied Soft Computing
Journal (2019) 105942, https://doi.org/10.1016/j.asoc.2019.105942.