International Journal of Technology Engineering rts Mathematics Science

Vol. 4, No. 1, June2024, pp. 01~06

ISSN: 2583-1224,
DOI: 10.11591/eei.v9i3.xxxx

Enhancing GPS Positioning Accuracy Using Machine Learning Regression

Durvesh Baharwal, 2Sanhita Dere, 3Ashwini Swami
1,2,3
Artificial Intelligence and Data Science Department, All India Shri Shivaji Memorial Society’s Institute of
Information Technology, Maharashtra, India
Corresponding Author: Durvesh Baharwal Email ID: [email protected]

Article Information ABSTRACT

Addressing the inherent challenges of Global Navigation Satellite Systems
(GNSS), this research project introduces an innovative approach by
combining Least Square estimation with the Time-Differenced Pseudo
Range method to enhance Position, Velocity, and Time (PVT)
determination. Enhancing GPS Positioning Accuracy Using Machine
Learning Regression aims to improve navigation systems, particularly in
Article history: challenging environments, through the integration of machine learning
Received May, 2024 within the framework of Least Square estimation. The methodology involves
a systematic integration process, showcasing the unique amalgamation of
Revised June,2024
traditional GNSS techniques with intelligent learning through Least Square
Accepted June,2024
estimation. The outcomes reveal significant improvements in navigation
accuracy, with Random Forest Regression emerging as the most effective
algorithm among those explored, maintaining its lead with the lowest MAE
of around 0.000122. Haversine distance is employed as a crucial metric for
quantitative evaluation. The project's practical implications extend to
mitigating delays and errors associated with GNSS, such as atmospheric
delays and multipath effects. The results underscore the transformative
impact of machine learning algorithms in refining GPS positioning accuracy
and set a new benchmark for precision in geospatial analysis and positioning
systems. The research concludes by highlighting the project's uniqueness,
practical applicability, and real-world adaptability—a tangible solution to
persistent challenges in satellite-based navigation.

KEYWORDS: GNSS (Global Navigation Satellite Systems), SLS (Simple

Least Square), WLS (Weight Least Square), Time-Differenced Pseudo Range
(TDP), PVT (Position Velocity Time), Machine Learning Regression,
Random Forest Regression, Navigation Accuracy.

1. INTRODUCTION of GNSS/INS integration strategies [1]. Zawislak et al.

‘Global Navigation Satellite System (GNSS)’ research (2022) significantly advance the exploration of
witnesses’ remarkable strides in recent years, with a machine learning applications in GNSS, focusing on
focus on elevating positioning accuracy and addressing multipath detection. Their unsupervised domain
challenges across diverse environments and adaptation approach, presented at the ION-GNSS+
applications. This literature review provides a Conference 2022, markedly enhances multipath
comprehensive overview of key contributions in this detection accuracy, showcasing potential for improving
dynamic landscape, encompassing the integration of GNSS positioning accuracy, especially in complex
technologies, innovative algorithms, and machine urban environments [2]. Maghdid et al. (2021) address
learning applications. challenges of accuracy and cost-efficiency in GNSS-
Boguspayev et al. (2023) offer a foundational based positioning through optimization approaches.
exploration of GNSS/INS integration techniques, Their survey emphasizes the integration of diverse
emphasizing the synergies between satellite navigation technologies and provides valuable insights into the
and inertial systems. Categorizing integration impact of optimization algorithms on accuracy
approaches, including loosely coupled, tightly coupled, improvement, time-to-fix reduction, and cost-
and deeply coupled methods, their work serves as a effectiveness [3]. Ji et al. (2022) contribute to the field
valuable reference, providing a nuanced understanding by evaluating GNSS-based velocity estimation

1
algorithms, addressing the critical need for precise ground truth records, satellite positions, and corrected
velocity determination in applications such as pseudoranges, followed by meticulous preprocessing
autonomous navigation and geodesy [4]. to ensure data integrity. Feature engineering extracts
Complementing these efforts, Kong (2021) introduces crucial features like corrected pseudoranges and
an innovative satellite positioning method employing satellite positions.
the Total Least Squares (TLS) algorithm, enhancing In the realm of traditional and improved position
accuracy and efficiency in satellite positioning [5]. estimation, both Traditional Least Square Estimation
Building upon these foundational studies, this research and Weighted Least Square Estimation are employed.
explores the application of advanced machine learning Traditional Least Square Estimation establishes a
regression techniques. The methods include ‘Support baseline, while Weighted Least Square Estimation
Vector Regression (SVR)’, XGBoost, ‘Decision Tree enhances accuracy by incorporating weightings to
prioritize more reliable measurements and mitigate the
Regression (DTR)’, and ‘Random Forest Regression
impact of outliers.
(RFR)’. Despite facing challenges in applying
The innovation lies in the integration of machine
Gaussian Process Regression (GPR) due to
learning regression techniques. ‘Support Vector
computational limitations, the study aims to contribute
Regression (SVR)’, XGBoost, ‘Decision Tree
to the ongoing development and improvement of
Regression (DTR)’, and ‘Random Forest Regression
GNSS-based navigation and positioning systems. (RFR)’ are employed for improved position
This research integrates insights from the literature estimation. SVR optimizes a hyperplane to minimize
survey to address the crucial aspects of precision and errors, XGBoost employs an ensemble of decision
reliability in GNSS systems. The application of trees, DTR leverages decision trees, and RFR
machine learning regression techniques opens new aggregates predictions from multiple decision trees.
avenues for achieving higher accuracy in diverse fields, Given below are the equations used for the respective
including autonomous vehicles, surveying, and algorithms that are used for the project.
geodetic applications.
4. SYSTEM ARCHITECTURE
2. PROPOSED SYSTEM This architecture represents an approach to GNSS
The proposed system architecture for enhancing GPS positioning that leverages machine learning to improve
positioning accuracy integrates traditional methods the accuracy and robustness of error mitigation
with advanced machine learning techniques. In the data techniques.
collection module, diverse GNSS data, including
ground truth records, satellite positions, and corrected
pseudo ranges, are acquired to represent various
environmental conditions. The data pre-processing
module employs robust techniques such as outlier
removal and noise reduction, ensuring the integrity of
the dataset. Feature engineering identifies critical
features like corrected pseudo ranges, forming the
foundation for both traditional Least Square Estimation
and advanced machine learning models. The system
incorporates ‘Support Vector Regression (SVR)’,
XGBoost, ‘Decision Tree Regression (DTR)’, and
‘Random Forest Regression (RFR)’ in the machine
learning integration module. Additionally, a ‘Gaussian
Process Regression (GPR)’ module attempts to model
non-linear relationships in the data. The quantitative
evaluation module employs metrics such as ‘mean
absolute error (MAE)’ and ‘mean squared error
(MSE)’ for a comprehensive analysis of model
accuracy. Results are then analyzed to understand the
strengths and limitations of each algorithm. The
proposed architecture aims to provide a robust and
adaptable solution for improving GPS accuracy in
diverse environments.

3. METHDOLOGY
This methodology integrates traditional techniques Figure 4.1: System Architecture
and advanced machine learning methodologies to
enhance GPS positioning accuracy. The initial phase
involves ingesting diverse raw GNSS data from

2
Where, GNSS OBS: This block likely refers to GNSS where X is the estimated position vector, ‘A’ is the
observations, which are the raw data received from design matrix, ‘W’ is the weight matrix, and ‘b’ is the
GNSS satellites. pseudo range residual vector.
Data Collection: This block collects GNSS
observations. 5.2 Weighted Least Square Estimation:
GPS/GNSS EPH: This block likely refers to ephemeris
data, which is broadcast by GNSS satellites and Weighted least squares estimation adjusts the
contains information about their orbits and positions. traditional method to account for measurement
Raw Pseudorange: This block indicates raw uncertainties. The estimation is given by:
pseudorange measurements, which are estimates of the
X = (ATWA)-1 AT Wb (5.2)
distance between a GNSS receiver and a satellite,
Where W is a diagonal matrix of weights, allowing for
affected by factors like satellite clock errors and
the down-weighting of less reliable measurements.
propagation delays.
GNSS Position Here, the GNSS position is calculated
6. MACHINE LEARNING REGRESSION
based on the raw pseudorange measurements.
TECHNIQUES
Basic Estimations: This block performs basic
estimations of the user's position using the 6.1. Support Vector Regression (SVR):
pseudorange measurements. ‘Support Vector Regression’ aims to find a hyperplane
Position Calculation: This block refines the position that best represents the underlying mapping function
calculation by incorporating error mitigation between the input features and the output. The SVR
techniques. objective function for regression can be defined as
Errors and Effects Mitigation: This block addresses follows:
various errors that affect the accuracy of GNSS
positioning, including: Objective Function:
Clock Synchronization Errors: These errors arise from Min V, b, l, l* ½ ||V||2 + C ∑ ( li + li* ) (6.1)
differences between the GNSS receiver clock and the
Subject to:
satellite clocks.
Yi – vT ϕ (Xi) – b ≤ € + li (6.2)
Ionospheric Errors: The ionosphere delays GNSS
vT ϕ (Xi) + b - Yi ≤ € + li* (6.3)
signals, affecting positioning accuracy.
li , li* ≥ 0 (6.4)
Tropospheric Errors: The troposphere, the Earth's
Here, ‘ϕ (xi) represents the feature mapping of the
lower atmosphere, also delays GNSS signals.
input data Xi’, ‘w is the weight vector’, ‘b is the bias
Relativistic Effects: Relativistic effects, based on
term’, ‘C is the regularization parameter’, and ϵ
Einstein's theory of relativity, can introduce minor
controls the width of the ϵ -insensitive tube.
errors in GNSS positioning.
Weighted Least Squares Estimator: This statistical
6.2. XGBoost:
method is likely used to refine the position calculation XGBoost is an ‘ensemble learning’ method that
by accounting for the precision of different combines the predictions from multiple decision trees.
measurements. The objective function for XGBoost regression can be
Machine Learning Layer: This block incorporates expressed as a sum of a loss term and regularization
machine learning algorithms to potentially improve the terms:
accuracy and robustness of error mitigation, especially
Objective Function:
regarding complex error sources that are difficult to
Obj = ∑ ( Yi + Yi* ) + ∑ ( fk ) (6.5)
model mathematically. The specific machine learning
Where ‘L (Yi + Yi* )’ is the loss term measuring the
algorithms mentioned include Support Vector
difference between the ‘Yi’ predicted and ‘Yi*’ actual
Regression (SVR), Extreme Gradient Boosting
values. ‘ (fk)’ is the regularization term for each tree.
(XGBoost), Decision Tree Regression (DTR),
Random Forest Regression (RFR). These algorithms
are likely trained on datasets containing GNSS 6.3. Decision Tree Regression (DTR):
‘Decision Tree Regression’ builds a tree structure to
observations, position information, and various error
predict the output value for a given set of input
sources.
features. The prediction for a new input \(x\) is
Output: The output of the system is a corrected GNSS
obtained by traversing the tree from the root to a leaf
position that is more accurate than the basic estimates
and outputting the average of the training target values
obtained using only pseudorange measurements.
in that leaf.
Prediction for Decision Tree Regression:
5. TRADITIONAL AND IMPROVED
Y*(X) – 1/Nk ∑i=leaf(X)Yi (6.6)
POSITION ESTIMATION Where ‘Y*(X)’ is predicted output for input X,
5.1 Traditional Least Square Estimation:
‘leaf(X)’ denotes the leaf node that input X falls into
X = (AT WA)-1 AT Wb (5.1) and ‘Nk’ is the number of training samples in leaf k.

6.4. Random Forest Regression (RFR):

3
‘Random Forest Regression’ aggregates predictions Random Forest Regression (RFR): 4.48 meters
from multiple decision trees, providing improved Support Vector Regression (SVR): 103.47 meters
generalization and robustness. The prediction for a Decision Tree Regression (DTR): 46.55 meters
new input x is obtained by averaging predictions from XGBoost: 947.25 meters
all the trees in the forest.

Prediction for Random Forest Regression:

Y*RFR(X) = 1/T ∑ Y*t(X) (6.7)
*
Where ‘Y RFR(X)’ is the predicted output for input X
using the Random Forest, Y*t(X) is the predicted
output for input X from tree t and T is the total number
of trees in the Random Forest.

6.5. Gaussian Process Regression (GPR):

The ‘Gaussian Process Regression’ model predicts
the corrected pseudo ranges (y) based on the satellite Figure 7.1: Plot for SVR, DTR, RFR and Actual
positions (X) and potentially other relevant features: Positions on Map
( ) (6.8)
where f (X) is the underlying function modeled by the When comparing the predicted positions of Support
Gaussian process, and ‘ε’ represents the ‘observation Vector Regression (SVR), Decision Tree Regression
noise’. (DTR), and Random Forest Regression (RFR) to the
The Gaussian Process model can be defined as: actual coordinates, distinct differences in accuracy are
f (X)∼ G P (μ (X), k (X, X’)) observed. The SVR model predicts coordinates with a
(6.9) deviation of 103.47 meters from the actual position,
where ‘μ (X)’ is the mean function and ‘k (X, X’)’ is indicating a moderate level of accuracy. On the other
the kernel function. hand, the DTR model provides a slightly better
For example, using the ‘Radial Basis Function (RBF) prediction with a distance of 46.55 meters from the
kernel’: actual coordinates. The RFR model, however,
– demonstrates the highest accuracy among the three,
( ) ( ) (6.10) with a minimal deviation of 4.48 meters. These results
These equations outline the fundamental principles highlight the superior performance of the RFR model
behind each regression algorithm, capturing the in predicting geographical positions, while both SVR
optimization objectives and prediction mechanisms. and DTR, although less accurate, still provide
reasonably close approximations to the actual values.
7. RESULTS AND DISCUSSION

The table provided represents the latitude and

longitude coordinates predicted by different machine
learning models compared to the actual coordinates.
To determine which model works the best and which
works the worst, we need to calculate the distance
between the actual coordinates and the predicted
coordinates for each model.

Table 7.1: Position Comparison Figure 7.2: Plot for XGBoost and Actual Positions on
Map
ML Model Latitude Longitude
When comparing the predicted position of the
Actual values 37.46758689 -122.1523673 XGBoost model to the actual coordinates, it is evident
Random Forest that XGBoost exhibits significant deviation, with a
37.46745277 -122.1527937 distance of 947.25 meters from the actual values. This
Regression (RFR)
large discrepancy highlights the model's lower
Support Vector
37.468082 -122.15336 accuracy in predicting geographical positions
Regression (SVR)
compared to other models such as Random Forest
Decision Tree
37.46716829 -122.1523673 Regression (RFR), Support Vector Regression (SVR),
Regression (DTR)
and Decision Tree Regression (DTR). The substantial
XGBoost 37.46513822 -122.1626473
error suggests that XGBoost may not be well-suited
for tasks requiring precise location predictions in this
Based on the distances calculated between the actual
specific context. Thus, while XGBoost is a powerful
coordinates and the predicted coordinates by each
and versatile model in many scenarios, its performance
model, here are the results:
4
in this case indicates room for improvement or the
need for further fine-tuning to achieve better accuracy
in predicting geographical coordinates.

Table 7.2: ‘Mean Absolute Error (MAE)’

Name of Mean Mean Mean
the Absolute Absolute Absolute
Machine Error Error Error Figure 7.1: Latitude Prediction ‘Mean Absolute Error
Learning (MAE) (MAE) (MAE) (MAE)’ comparison for respective Regression
Model Latitude Longitude Height Models
Random
Forest
0.000081 0.000122 0.063725
Regression
(RFR)
Support
Vector
0.000014 0.000052 2.361806
Regression
(SVR)
XGBoost 0.000366 0.000528 0.129267 Figure 7.1: Longitude Prediction ‘Mean Absolute
Decision Error (MAE)’ comparison for respective Regression
Tree
0.000179 0.000253 0.090862 Models
Regression
(DTR)
In the comprehensive evaluation of various regression
algorithms aimed at augmenting GPS positioning
accuracy, the Mean Absolute Error (MAE) values play
a pivotal role in gauging their effectiveness. In the
realm of latitude prediction, Random Forest
Regression (RFR) emerged as the standout performer,
boasting the lowest MAE of approximately 0.000081. Figure 7.3: Height Prediction ‘Mean Absolute Error
This signifies RFR's exceptional precision in (MAE)’ comparison for respective Regression Models
estimating the north-south position of GPS
coordinates. Support Vector Regression (SVR) closely
followed suit, exhibiting remarkable accuracy with a This analysis indicates that the Random Forest
MAE of about 0.000014, particularly excelling in the Regression (RFR) model provides the most accurate
latitude dimension. The prowess of RFR persisted in predictions for the given coordinates, while the
longitude prediction, where it maintained its lead with XGBoost model provides the least accurate
the lowest MAE of around 0.000122, showcasing its predictions.
superior performance in capturing east-west
coordinates. In this aspect, XGBoost and Decision 8. CONCLUSION
Tree Regression (DTR) displayed slightly higher
The project aims to systematically enhance GPS
MAEs, indicating marginally reduced accuracy
compared to RFR. When considering height positioning accuracy by integrating traditional
prediction, RFR continued to outshine other methods with machine learning regression techniques.
algorithms, achieving the lowest MAE of The methodology involves meticulous data
approximately 0.063725. This reinforces its robust preprocessing and feature engineering, ensuring a
performance in accurately determining the vertical robust foundation for analysis. Traditional Least
position of GPS coordinates. Although XGBoost Square Estimation establishes a baseline for position
exhibited a higher MAE of about 0.129267 in height estimation, while Weighted Least Square Estimation
estimation, it demonstrated competitive results, enhances accuracy by considering measurement
underlining its effectiveness. In summation, Random
uncertainties. The innovation lies in the application of
Forest Regression consistently proved to be the most
effective algorithm across all dimensions, affirming its machine learning regression techniques, including
efficacy in significantly enhancing GPS positioning Support Vector Regression, XGBoost, Decision Tree
accuracy. Regression, and Random Forest Regression, to
improve position estimation. The research provides a
comprehensive evaluation of model performance using
mean absolute error and mean squared error metrics.

5
The integration of machine learning within the
framework of traditional techniques showcases the 9. REFERENCES
potential for achieving higher accuracy and reliability
in GPS positioning. Future work will explore further [1] Boguspayev, A., Zhilong, L., & Zhang, X. (2023).
Integration of GNSS/INS techniques: A survey. Journal of
optimization and real-world implementation, Navigation and Control, 10(2), 101-11
contributing to the ongoing development of GNSS-
based navigation and positioning systems.
[2] Zawislak, P., Borkowski, M., & Wrobel, J. (2022). [6] Zhipeng Wang, Xiaopeng Hou, Zhiqiang Dan & Kun
Enhancing GNSS Multipath Detection with Unsupervised Fang, "Adaptive Kalman filter based on integer ambiguity
Domain Adaptation. In Proceedings of the ION-GNSS+ validation in moving base RTK".
Conference 2022 (pp. 1234-1245).
[7] Min Li, Tianhe Xu, Yali Shi, Kai Wei, Xianming Fei,
[3] Maghdid, H. J., Lami, K. S., & Ghafoor, K. Z. (2021). Dixing Wang, 1"Adaptive Kalman Filter for Real-Time
Optimization approaches in GNSS-based positioning: A Precise Orbit Determination of Low Earth Orbit Satellites
review. GPS Solutions, 25(3), 89-102. Based on Pseudorange and Epoch-Differenced Carrier-Phase
[4] Ji, X., Zhiqiang, W., & Lu, Y. (2022). A review of Measurements".
GNSS-based velocity estimation algorithms. Journal of
Navigation, 75(1), 1-19. [8] Haibo Ge, Bofeng, Icon,Song, Jia,Liangwei Nie, Tianhao
Wu, Zhe Yang, Jingzhe Shang, Yanning Zheng & Maorong
[5] Kong, L. (2021). An improved satellite positioning Ge, "LEO Enhanced Global Navigation Satellite System
method based on the total least squares algorithm. GPS (LeGNSS): progress, opportunities, and challenges".
World, 32(4), 54-63.
[9] Eugenio Realini, "goGPS free and constrained relative
kinematic positioning with low cost receivers"