**Abstract**

This paper presents a machine learning-based approach for real-time kinematic (RTK)
positioning using global navigation satellite systems (GNSS). RTK positioning offers highly
accurate position solutions in the centimeter range, making it valuable for various applications
such as autonomous vehicles, mobile robotics, and geospatial data collection. The proposed
system leverages machine learning algorithms to enhance the robustness and accuracy of RTK
positioning.

The proposed system extracts features from GNSS data that are relevant to RTK positioning.
Several machine learning algorithms are considered for the task of RTK positioning. The
proposed system is deployed in a real-time system and can achieve centimeter-level positioning
accuracy.

Introduction:

Real-time kinematic (RTK) positioning is a differential positioning method that utilizes signals
from global navigation satellite systems (GNSS) to provide highly accurate position solutions in
real-time. With nominal accuracy on the order of centimeters, RTK positioning has become
indispensable in applications such as autonomous vehicles, mobile robotics, driver-assist
technologies, and precise geospatial data collection. However, challenges still exist in achieving
robust and reliable RTK solutions, especially in diverse and dynamic environments.

In recent years, machine learning techniques have shown great promise in improving the
accuracy and robustness of positioning systems. By leveraging the power of data-driven
models, machine learning algorithms can capture complex spatial and temporal relationships
inherent in GNSS data, leading to enhanced RTK positioning performance. This research
explores the application of machine learning algorithms to develop a real-time kinematic
positioning system for GNSS receivers.

The primary objective of this study is to leverage machine learning algorithms to address the
limitations and challenges faced by traditional RTK positioning methods. By incorporating these
algorithms, we aim to improve the accuracy, reliability, and generalizability of RTK solutions
across various environmental conditions and receiver configurations.

To achieve this objective, the research follows a systematic approach. First, a comprehensive
understanding of the principles and techniques involved in RTK positioning and GNSS systems
is presented. This includes an overview of the underlying GNSS technology, the differential
positioning concept, and the advantages and limitations of RTK methods.

Next, a dataset is collected, comprising GNSS data from receivers and corresponding ground
truth RTK position measurements. The GNSS data includes essential measurements such as
satellite positions, signal strengths, and receiver parameters. This dataset forms the foundation
for training and evaluating the machine learning models.

Feature engineering techniques are employed to extract relevant information from the GNSS
data. These features capture spatial and temporal patterns that are crucial for accurate
positioning. By carefully selecting and transforming the features, we aim to provide the machine
learning models with comprehensive input representations that facilitate robust RTK positioning.
Several machine learning algorithms suitable for regression tasks are evaluated for their
efficacy in RTK positioning. Recurrent neural networks (RNNs), convolutional neural networks
(CNNs), support vector machines (SVMs), and random forests are considered as potential
models. These algorithms are trained on the labeled dataset, enabling them to learn the
complex relationships between the extracted features and the corresponding RTK position
measurements.

The performance of the trained machine learning models is thoroughly evaluated using
appropriate metrics, such as centimeter-level positioning accuracy and error measures. The
models showing the best performance are selected for deployment in a real-time system,
capable of processing GNSS data and providing accurate RTK positioning outputs with low
latency.

Continuous monitoring and improvement play a vital role in ensuring the reliability and
generalizability of the developed system. The performance of the machine learning-based RTK
positioning system is rigorously tested in real-world scenarios, considering diverse
environments, receiver types, and motion dynamics.

By combining the strengths of GNSS technology and machine learning algorithms, this research
aims to advance the field of real-time kinematic positioning. The outcomes of this study have the
potential to contribute to the development of more accurate and robust GNSS-based positioning
systems, opening new possibilities for applications in autonomous navigation, robotics, and
geospatial data collection.

Overall, this research endeavors to bridge the gap between GNSS technology and machine
learning, enabling the realization of precise and reliable real-time kinematic positioning solutions
for various practical applications.

Machine Learning Algorithm:

In this study, machine learning algorithms were employed to enhance the real-time kinematic
(RTK) positioning system using global navigation satellite systems (GNSS). While the specific
algorithms used in the original thesis excerpt are not mentioned, we will describe some
commonly used algorithms that could be suitable for RTK positioning:

1. Recurrent Neural Networks (RNNs): RNNs are a class of deep learning models capable of
processing sequential data. They are well-suited for tasks involving temporal dependencies,
making them a suitable choice for RTK positioning. RNNs can capture the sequential patterns in
GNSS data and learn the relationships between historical measurements and the corresponding
RTK positions. Variants of RNNs, such as Long Short-Term Memory (LSTM) and Gated
Recurrent Units (GRUs), are particularly effective in capturing long-term dependencies.

2. Convolutional Neural Networks (CNNs): CNNs are primarily used for image analysis, but they
can also be adapted for sequential data processing. In the context of RTK positioning, 1D CNNs
can be utilized to capture spatial patterns or analyze specific characteristics of the GNSS
signals. By applying filters and pooling operations, CNNs can extract relevant features from the
input data, enabling the model to learn and predict accurate RTK positions.
3. Support Vector Machines (SVMs): SVMs are a classical machine learning algorithm
commonly used for regression tasks. They can effectively handle nonlinear relationships in data
and generalize well to unseen samples. SVMs can be trained to predict RTK positions based on
the extracted features from the GNSS data. With proper feature engineering and parameter
tuning, SVMs can offer accurate and robust RTK positioning solutions.

4. Random Forests: Random Forests are an ensemble learning method that combines multiple
decision trees. They are known for their ability to handle high-dimensional data and capture
complex relationships. Random Forests can be used for RTK positioning by training an
ensemble of decision trees on the extracted features. The model can then provide predictions
based on the collective knowledge of the trees, resulting in accurate and reliable RTK positions.

Problem Statement:

Real-time kinematic (RTK) positioning using global navigation satellite systems (GNSS) offers
highly accurate position solutions with centimeter-level precision. However, there are challenges
and limitations associated with traditional RTK methods, such as susceptibility to environmental
conditions, signal interference, and receiver limitations. These challenges can lead to reduced
accuracy, limited availability, and decreased reliability of the positioning solutions.

In addition, the cost of high-end RTK receivers can be prohibitive for many applications, making
it necessary to explore the potential of low-cost receivers. However, the performance and
reliability of these low-cost RTK receivers need to be thoroughly assessed and compared to
ensure their suitability for various real-world applications.

Moreover, despite the advancements in RTK positioning, there is a need to enhance the
robustness and reliability of the solutions in dynamic environments. Unpredictable motion,
changes in signal quality, and varying environmental conditions pose challenges that need to be
addressed to ensure accurate and consistent RTK positioning performance.

Furthermore, the forecasting aspect of RTK positioning systems using machine learning
techniques requires exploration and improvement. The ability to predict and estimate RTK
positions in real-time based on historical GNSS data is crucial for achieving higher accuracy and
robustness. Therefore, the development of effective and reliable forecasting methods using
machine learning algorithms needs to be investigated and refined.

Overall, the problem statement revolves around the need to address the limitations of traditional
RTK positioning methods, assess the performance of low-cost RTK receivers, improve the
robustness of RTK solutions in dynamic environments, and enhance the forecasting capabilities
using machine learning algorithms.

By addressing these challenges, it is possible to develop a machine learning-based RTK

positioning system that offers accurate, reliable, and cost-effective solutions for applications
such as autonomous vehicles, mobile robotics, and precise geospatial data collection.

Motivation and Needs:

The motivation behind developing a machine learning-based real-time kinematic (RTK)
positioning system stems from the following key factors:

1. Enhanced Positioning Accuracy: Traditional RTK positioning methods have limitations that
can result in reduced accuracy, especially in challenging environments. By integrating machine
learning algorithms, it is possible to exploit the inherent patterns and relationships in the GNSS
data to achieve higher accuracy. Improved positioning accuracy is crucial for applications such
as autonomous vehicles, mobile robotics, and precise geospatial data collection, where
centimeter-level precision is necessary.

2. Increased Robustness and Reliability: RTK positioning systems face challenges in dynamic
environments, including signal interference, multipath effects, and varying atmospheric
conditions. These factors can impact the robustness and reliability of the positioning solutions.
By leveraging machine learning algorithms, it becomes possible to model and account for these
environmental variations, resulting in more robust and reliable RTK solutions. This is vital for
ensuring the safety and effectiveness of applications in real-world scenarios.

3. Cost-Effectiveness: High-end RTK receivers can be expensive, making them inaccessible for
many applications that require precise positioning. By exploring the capabilities of low-cost RTK
receivers and assessing their performance, it becomes possible to provide cost-effective
alternatives without compromising on accuracy or reliability. This is particularly important for
widespread adoption of RTK positioning in industries where cost constraints are a significant
factor.

4. Real-Time Predictive Capabilities: Incorporating machine learning algorithms enables the

development of forecasting methods for RTK positioning. Real-time prediction of RTK positions
based on historical GNSS data enhances the system's ability to anticipate and adapt to
changes, leading to improved accuracy and robustness. Predictive capabilities are essential in
dynamic scenarios where the instantaneous knowledge of accurate positions is critical for
decision-making processes.

5. Practical Applications and Impact: The development of a machine learning-based RTK

positioning system addresses the needs of various industries and applications. Autonomy in
vehicles, precision in robotics, efficient data collection in geospatial applications, and driver-
assist technologies all rely on highly accurate and reliable positioning. By meeting these needs,
the research has the potential to revolutionize these fields, improving safety, efficiency, and
performance.

Overall, the motivation behind this research lies in the desire to overcome the limitations of
traditional RTK positioning methods, achieve higher accuracy and robustness, reduce costs
through low-cost alternatives, provide real-time predictive capabilities, and cater to the needs of
practical applications. By addressing these motivations and needs, the study aims to contribute
to advancements in the field of RTK positioning and its application in various industries and
domains.

Challenges:

The development and implementation of a machine learning-based real-time kinematic (RTK)

positioning system using global navigation satellite systems (GNSS) entail several challenges:
1. Data Quality and Availability: Ensuring the availability of high-quality GNSS data is crucial for
accurate and reliable RTK positioning. Factors such as signal obstructions, multipath effects,
and atmospheric disturbances can introduce noise and errors into the data. Managing and
mitigating these issues to maintain data quality and availability pose a significant challenge.

2. Dataset Size and Diversity: Constructing a comprehensive and representative dataset for
training the machine learning models can be challenging. It requires collecting a large and
diverse dataset that covers a wide range of environmental conditions, receiver types, and
motion dynamics. Obtaining such a dataset may involve extensive fieldwork and collaboration
with multiple sources, which can be time-consuming and resource-intensive.

3. Feature Engineering: Extracting meaningful and relevant features from the GNSS data is
crucial for the success of the machine learning models. Determining the most informative
features that capture the spatial and temporal patterns while filtering out irrelevant noise can be
challenging. It requires domain expertise and careful analysis of the data to select and engineer
appropriate features for optimal model performance.

4. Model Selection and Configuration: Choosing the most suitable machine learning algorithms
for RTK positioning can be challenging. There are various models available, each with its own
strengths and limitations. Determining the optimal architecture, hyperparameters, and training
configurations that yield the best performance requires experimentation, validation, and
expertise in machine learning.

5. Real-Time Processing: Implementing the machine learning-based RTK positioning system to

operate in real-time presents challenges. The system must process incoming GNSS data
rapidly, perform feature extraction, and make predictions or estimates within tight latency
constraints. Efficient algorithms, hardware acceleration, and optimization techniques are
necessary to achieve low-latency processing while maintaining accuracy.

6. Generalization and Adaptability: Ensuring that the machine learning models generalize well to
unseen data and different environments is crucial for practical applications. The models need to
adapt and provide accurate RTK positions in diverse scenarios, including varying environmental
conditions, receiver configurations, and motion dynamics. Achieving generalization and
adaptability requires careful training, validation, and testing on representative datasets.

7. System Integration and Deployment: Integrating the machine learning-based RTK positioning
system into existing infrastructure or applications can be challenging. Compatibility with different
GNSS receivers, data formats, and communication protocols needs to be considered. The
system should be designed for easy deployment, maintenance, and scalability to accommodate
diverse real-world scenarios.

8. Ethical and Legal Considerations: Collecting and processing GNSS data raises privacy
concerns and ethical considerations. It is crucial to handle data in compliance with privacy
regulations and obtain necessary permissions when using real-world data. Ensuring the
responsible and ethical use of GNSS data throughout the research process is paramount.

Challenges:
The development and implementation of a machine learning-based real-time kinematic (RTK)
positioning system using global navigation satellite systems (GNSS) entail several challenges:

1. Data Quality and Availability: Ensuring the availability of high-quality GNSS data is crucial for
accurate and reliable RTK positioning. Factors such as signal obstructions, multipath effects,
and atmospheric disturbances can introduce noise and errors into the data. Managing and
mitigating these issues to maintain data quality and availability pose a significant challenge.

2. Dataset Size and Diversity: Constructing a comprehensive and representative dataset for
training the machine learning models can be challenging. It requires collecting a large and
diverse dataset that covers a wide range of environmental conditions, receiver types, and
motion dynamics. Obtaining such a dataset may involve extensive fieldwork and collaboration
with multiple sources, which can be time-consuming and resource-intensive.

3. Feature Engineering: Extracting meaningful and relevant features from the GNSS data is
crucial for the success of the machine learning models. Determining the most informative
features that capture the spatial and temporal patterns while filtering out irrelevant noise can be
challenging. It requires domain expertise and careful analysis of the data to select and engineer
appropriate features for optimal model performance.

4. Model Selection and Configuration: Choosing the most suitable machine learning algorithms
for RTK positioning can be challenging. There are various models available, each with its own
strengths and limitations. Determining the optimal architecture, hyperparameters, and training
configurations that yield the best performance requires experimentation, validation, and
expertise in machine learning.

5. Real-Time Processing: Implementing the machine learning-based RTK positioning system to

operate in real-time presents challenges. The system must process incoming GNSS data
rapidly, perform feature extraction, and make predictions or estimates within tight latency
constraints. Efficient algorithms, hardware acceleration, and optimization techniques are
necessary to achieve low-latency processing while maintaining accuracy.

6. Generalization and Adaptability: Ensuring that the machine learning models generalize well to
unseen data and different environments is crucial for practical applications. The models need to
adapt and provide accurate RTK positions in diverse scenarios, including varying environmental
conditions, receiver configurations, and motion dynamics. Achieving generalization and
adaptability requires careful training, validation, and testing on representative datasets.

7. System Integration and Deployment: Integrating the machine learning-based RTK positioning
system into existing infrastructure or applications can be challenging. Compatibility with different
GNSS receivers, data formats, and communication protocols needs to be considered. The
system should be designed for easy deployment, maintenance, and scalability to accommodate
diverse real-world scenarios.

8. Ethical and Legal Considerations: Collecting and processing GNSS data raises privacy
concerns and ethical considerations. It is crucial to handle data in compliance with privacy
regulations and obtain necessary permissions when using real-world data. Ensuring the
responsible and ethical use of GNSS data throughout the research process is paramount.
Addressing these challenges requires a multidisciplinary approach, combining expertise in
GNSS technology, machine learning, signal processing, and system integration. Thorough
experimentation, validation, and continuous improvement are essential to overcome these
challenges and develop a robust and practical machine learning-based RTK positioning system.