This is a self-archived – parallel published version of this article in the publication archive of the University of Vaasa. It might differ from the original. Machine Learning Utilization in GNSS: Use Cases, Challenges and Future Applications Author(s): Siemuri, Akpojoto; Kuusniemi, Heidi; Elmusrati, Mohammed S.; Välisuo, Petri; Shamsuzzoha, Ahm Title: Machine Learning Utilization in GNSS: Use Cases, Challenges and Future Applications Year: 2021 Version: Accepted version Copyright ©2021 IEEE. Personal use of this material is permitted. Permission from IEEE must be obtained for all other uses, in any current or future media, including reprinting/republishing this material for advertising or promotional purposes, creating new collective works, for resale or redistribution to servers or lists, or reuse of any copyrighted component of this work in other works. Please cite the original version: Siemuri, A., Kuusniemi, H., Elmusrati, M. S., Välisuo, P. & Shamsuzzoha, A. (2021). Machine Learning Utilization in GNSS: Use Cases, Challenges and Future Applications. In: 2021 International Conference on Localization and GNSS (ICL- GNSS), 20843286. United States: IEEE. https://doi.org/10.1109/ICL-GNSS51451.2021.9452295 Machine Learning Utilization in GNSS—UseCases, Challenges and Future Applications Akpojoto Siemuri; Petri Välisuo; Heidi Kuusniemi; Mohammed S. Elmusrati; Ahm Shamsuzzoha All authors are from the University of Vaasa. Abstract— The algorithms and models of traditional global navigation satellite systems (GNSSs) perform very well in terms of theavailability and accuracy of positioning, navigation and timing (PNT) under good signal conditions. Research is still ongoing toimprove their robustness and performance in less than optimal signal environments. A growing interest in the study of machinelearning (ML) and the potential for its application in many fields has also led to several types of research on its utilization in GNSSs.In the field of GNSSs, ML is changing the ways that navigation problems are prevented and resolved, and it is taking on a significantrole in advancing PNT technologies for the future. We illustrate this point by reviewing how ML can enhance GNSS performanceand usability and also discuss areas of GNSSs in which ML algorithms have been applied. We also highlight the commonlyimplemented ML algorithms and compare their performance when used in similar GNSS use cases. In addition, the challenges andrisks of the utilization of ML techniques in GNSSs are discussed. Insight is given into prospective areas in GNSSs in which ML canbe applied for increased performance, accuracy and robustness, thereby providing fertile ground for novel research. Keywords— Machine Learning (ML), Deep Learning (DL), Global Navigation Satellite Systems (GNSSs), PNT technologies, GNSSperformance. I. INTRODUCTION In global navigation satellite systems (GNSSs), there are several sources of errors for satellite-based positioning that affect accuracy anddegrade receiver performance, namely, the ionospheric effect, multipath propagation, spoofing and jamming interferences and GNSS-deniedenvironments. Other sources of error in GNSSs include receiver noise and resolution, satellite clock errors and hardware biases [1]. Therefore,it is interesting to see how machine learning (ML) has and can be applied in such cases.Machine-learning techniques are well suited for real-life problems and are tolerant to data that are imprecise, partially incorrect oruncertain [2]. ML is a very powerful technique in handling time-series data, and it can learn time-dependent patterns across multiple models.Therefore, ML can be found useful in the discovery of unknown and hidden information in GNSS data. The nearly limitless quantity of availabledata, affordable data storage and the growth of less expensive and more powerful processing has propelled the growth of ML. Compared tostatistical methods, ML methods enable us to identify tricky dependencies in datasets for which exploratory analysis has not enabled the properdetermination of the shape of the underlying model. The aim in ML is not to provide an explicit formula for the distribution of the data; rather,an algorithm is trained to detect the relationships among the features of the data on its own, directly from the data. Learning methods makes itpossible to avoid the assumptions involved in many statistical methodologies. A demonstration of the concept of training using a machine-learning approach is shown in Fig. 1. It shows an ML model trained to classify a GNSS signal as line of sight (LOS)/multipath/non-line of sight(NLOS). This paper presents some major applications in GNSSs in which ML has been utilized to provide a novel solution or a new service. Thedifferent ML algorithms, methods and solutions used in the primary literature are compared to show how they are used. The rest of the paper isorganized as follows: Section II presents the review process and the research questions that are discussed. In Sections III through V, the researchquestions used are discussed, while Section VI discusses the challenges and risks of using ML algorithms in GNSSs. Section VII presents theconclusions and future directions. II. FEASIBILITY ASSESSMENT To facilitate the use of ML algorithms in GNSSs, it is necessary to review their performance and usage in the extant literature. This reviewbegan with a systematic search for applications of ML in GNSS use cases of digital libraries, leading to the retrieval of 182 papers. A total of47 of these papers were deemed relevant and required full-text screening for detailed information on the application of ML in a GNSS. Thesearch phrase used to identify the primary studies was “GNSS AND (machine learning OR deep learning OR artificial intelligence OR randomforest OR support vector machine OR decision tree OR neural network OR regression)”. The four electronic databases used for the searchwere IEEE Xplore, ScienceDirect, Google Scholar and the Web of Science. We restricted the search to articles published from 2000 to 2020. Fig. 1. Demonstration of the concept of training using amachine-learning approach [3]. Two other papers ([43 and [50]) were added based on recommendations from reviewers. The following research questions (RQs) wereidentified to guide the review: RQ1. To which GNSS use cases have ML algorithms been applied and in what are other prospective areas of GNSSs could ML algorithmsbe utilized? RQ2. Which ML algorithms are utilized in GNSSs and are there ML algorithms that significantly outperform the others? RQ3. What are the challenges and risks of using ML algorithms in GNSSs? These research questions are discussed in Sections III through V, respectively. III. MACHINE LEARNING UTILIZATION IN GNSS USE CASES The use of ML in GNSSs is promising, as it has proven useful in various studies. Here, we discuss the ML techniques implemented in theprimary studies. The ML techniques described include decision trees (DTs), gradient-boosting decision trees (GBDTs), logistic regression (LR),long short-term memory (LSTM), naïve Bayes (NB), Gaussian process regression (GPR), deep learning (DL) models (neural networks, or NN),extreme learning machines (ELMs), support vector machines (SVM) etc. The algorithms were used for classification, clustering, forecastingand anomaly detection depending on the GNSS application. ML cannot completely replace GNSS physical models, but the use of ML/DLtogether with traditional GNSS has been promising and has become more popular among researchers, as discussed in the following subsections. A. GNSS Signal Acquisition The signal acquisition process estimates if a particular satellite signal is present or not in the received signal, and it also gives a roughestimate of its associated code delay and Doppler frequency if present [4]. This acquisition process is implemented by all GNSSreceivers [5], [1]. This process is achieved through the evaluation of a cross-ambiguity function (CAF), usually in a discrete-time domain. TheCAF is a two-dimensional function that is related to the correlation between the received signal and local code for every possible delay/Dopplerpair. The CAF, which can be considered an image, has certain characteristics that can be used to distinguish the presence or absence of a signalfrom a specific satellite. This knowledge can be used to train a data-driven model, such as a multi-layer perceptron (MLP), which is a neuralnetwork architecture with moderate complexity that has been widely used in ML studies, or a convolution neural network (CNN), which is ableto capture complex non-linear phenomena, at the expense of much greater complexity when compared to MLPs [4]. B. Signal Detection and Classification Most of the solutions already in the market that implement multipath detection and mitigation are designed using stochastic modelling,spatial geometry modelling, advanced techniques in data processing, or new special hardware designs [6]; however, this is not the case for allreceivers, for example, low-cost receivers and GNSS-enabled smartphones. In the detection and classification of the characteristics of anacquired signal, the model needs to be able to accurately and reliably classify the LOS, multipath and NLOS signals. However, when there isan event outside the statistical models (too complex to be modelled with classical statistics and that does not fit the mathematical assumptionsused to develop the models), these solutions become ineffective. ML methods enable us to throw off assumptions required in the statisticalmethodology. In [7], [6], [8], [9] and [10], ML methods, such as logistic regression (LR), support vector machines (SVMs), naïve Bayes (NB),decision trees (DTs), CNNs and recurrent neural networks (RNNs), have been implemented. An artificial neural network (ANN) modelcapable of processing the structure of the autocorrelation function (ACF) was used for the detection of evil waveforms (EWF) in [11]. In [12]and [13], a robust gradient-boosting decision tree (GBDT) and DT-based classifier were employed, respectively, for GNSS signal receptionclassification that made use of the carrier-to-noise-density ratio (C/N0), pseudo-range residuals and satellite elevation angle as the inputfeatures to improve the performance of the signal classification at the receiver. C. Earth Observation and Monitoring GNSSs have been used in understanding the atmospheric effects of earthquakes by analysing the ionospheric behaviour before and afterthem through the determination of the total electron content (TEC) using GNSS data [14]. TEC is a representation of the electron density in thesignal trajectory between the satellite and the receiver on the earth’s surface. In [14], vertical total electron content (VTEC) data from theNational Oceanic and Atmosphere Administration (NOAA) along with ionospheric disturbance information were used for training an ANN forthe prediction and detection of earthquakes and the determination of their magnitude. Hurricane detection and prediction are based on thecategorization of the maximum wind speed that can be estimated from the reflected GNSS signals, but to be able to track the hurricaneefficiently, a large amount of measurement data is required. This is where a deep learning algorithm such as CNN is used, as seenin [15]. Similarly, in [16], a CNN model based on classification was used for sea ice detection, while a regression-based CNN was applied insea ice concentration (SIC) estimation from TechDemoSat-1 GNSS reflectometry delay-Doppler maps (DDMs). DL has also been useful inother cases, such as environmental remote sensing, which includes land cover mapping, aerosol monitoring and the prediction of agriculturalyield, snow cover, ocean colour parameters etc. [17]. D. GNSS Navigation and Precise Positioning Nowadays, GNSS is successfully implemented to achieve precise positioning in both indoor and outdoor environments [18], [19]. Suchprecise positioning is essential for safe operations [20], [21]. For indoor positioning, GNSS is applied in areas like warehouse management[22], automated manufacturing [23] etc. In the case of outdoor positioning, GNSS is also suitable for tracking and tracing in various sectors,such as intelligent transportation systems (ITSs), autonomous vehicles [24], supply chain and logistics [25], port operations [26] etc.Achieving accurate and precise positioning is achieved by minimizing errors in both indoor and outdoor positioning.ML has been applied in location-based services (LBS) with the aim of improving GNSS navigation and positioning in several scenarios. In[27], the least-squares support vector machine (LSSVM) technique was used to enhance the accuracy of Kalman filtering (KF). This enhancedmodel is called LSSVM-enhanced KF (LSSVM-KF). The KF relies on the quality of the observations as well as the dynamic model; therefore,LSSVM-KF adaptively estimates the dynamic modelling bias from the historical information and then uses the bias estimate to compensate thedynamic model. The dynamic model bias is treated as a time-variant ambiguous function by the algorithm, which is trained with the LSSVM[27]. Also, ML has been applied in ITSs to estimate the GNSS position error by aiding motion sensor units in providing a more accurate positionestimate during periods of outage or blockage of the GNSS signal. In [28], a DT and SVM were implemented to choose the best feature forclassifying the input data at every iteration and for the selection of features that were most relevant for the location error estimation. E. Indoor Navigation Scenario recognition is essential for seamless indoor localization and robust positioning in complex environments. In [29], the influenceof multi-constellation GNSS measurements on scenario recognition performance was studied using a hidden Markov model (HMM)algorithm. This work was based on the recognition of four scenarios (deep indoor, shallow indoor, semi-outdoor and open outdoor). Theresults from the scenario recognition showed a significant improvement with the increase in the number of constellations received by thesmartphones. However, when the user was switching between indoor and outdoor environments, the observed variables (such as the numberof visible satellites) did not respond to this change of environment immediately, as the change to the current scenario was rather slow. Thiscaused the result of scenario recognition to lag behind the true value, and it is a challenge to effectively shorten this transition delay duringscenario recognition. A scenario recognition algorithm based on an RNN, which showed better results in terms of accuracy, adaptability tonew environments and real-time performance, was implemented to meet this challenge. F. Ionospheric Scintillations and Tropospheric Wet Delay In a GNSS receiver, signal acquisition and tracking can be severely impacted by strong scintillation, resulting in GNSS performancedegradation in accuracy and continuity. It is difficult to predict and model scintillation because there are different reasons for the occurrenceof this phenomenon, for example, solar activity, magnetic storms, local electric fields, conductivity, wave interaction etc. [30]. A DT modelin [31] and an SVM in [30] are examples of ML algorithms employed for the detection of scintillation. The time delay of a GPS (L1 and L2)signal in the ionosphere is an example of propagation path delay, and it depends on the TEC of the atmospheric layer. This delay contributesto a potential source of error in time measurements and can produce an error range in tens of meters. In [32], Gaussian process regression(GPR) was used in comparison to an auto regressive moving average (ARMA) model and ANN model to implement the forecasting of low-latitude ionospheric conditions, and in [33], a single frequency correction algorithm based on a GNSS global ionosphere map (GIM) verticalTEC (vTEC) and fully connected neural networks (FC-NN) was used to estimate GNSS single-frequency ionospheric delay. [34] is anotherstudy on GNSS atmosphere delays in which ML has been employed. In [35], an ANN model was implemented to predict tropospheric wetdelay based on meteorological and GNSS data. G. GNSS Security—Spoofing and Jammer Attacks Mitigation against spoofing attacks has been achieved at the pseudorange level, with receivers employing receiver autonomous integritymonitoring (RAIM) by the discovery of an inconsistent set of five or more pseudoranges to allow the receiver to identify an unsophisticatedspoofer that broadcasts one or more false signals with no attempt to achieve believable consistency. However, in response to efforts to defendagainst spoofing, advanced forms of GNSS spoofing have been conceived, for example, a spoofing method called “in the wild”, which is anactual malicious spoofing attack [36]. Therefore, ML has been implemented to detect and defend against spoofing; for example, in [37], one-and two-hidden-layer neural networks with various numbers of hidden neurons were implemented to detect GPS spoofing signals by makinguse of different features, such as pseudo-range, Doppler shift and signal-to-noise ratio (SNR), to perform the classification of GPS signals.While [38] used an MLP neural network classifier trained by particle swarm optimization (PSO). Other ML algorithms used to detect jammingand spoofing include RNN based on long short-term memory (LSTM), classification SVM (C-SVM) with principal component analysis(PCA), and SC-SVM, as seen in [39], [40] and [41], respectively. In the case of GNSS jamming, in [42], SVM and CNN models were usedfor the detection and classification of the jammer signal. In [43], ML is used to protect users against Secure Code Estimation and Replay(SCER) attacks on Galileo OS-NMA making use of the analysis done on the features extracted from the victim’s Search Space. This MLtechnique is, however, complementary to the Navigation Message Authentication (NMA) techniques meaning that it will not offer anyguarantee the navigation message was not modified if the ML is used separately without the NMA technique. The features extraction used isbased on fitting the correlation peaks in the search space as 2D Gaussians and detecting radio frequency interferences (RFIs) during a timeanalysis window. The detection used in this paper is based on simply finding outliers with the assumption that an Automatic Gain Controller(AGC) is present. Using algorithms based on Decision Trees (DT, ADA Boost, and Random Forest) gave the best results when signal C/N0is above 30dBHz. Other algorithms implemented were Nearest Neighbour, and RBF SVM algorithm with the accuracy decreasing when thelocation of the peaks is introduced, particularly for high C/N0. While Linear SVM showed that as the C/N0 improves the results also improve,and give values similar to that of the Decision Trees when C/N0 is greater than 30 dBHz. H. GNSS/Inertial Navigation Systems (GNSS/INS) Integration Several technologies that are essential to reducing positioning errors are now being researched [28]. Nowadays, a GNSS is interfacedwith inertial navigation systems (INS) along with some filtering techniques with the objective of improving the overall positioning system[44]. Usually, a combination of an INS with various filters is used to calculate the accurate position. In such a situation, a GNSS outage ofshorter duration does not have much effect on estimating accurate positioning. However, if the outage becomes longer, then the accuracy ofthe positioning decreases significantly [45]. Therefore, it is a common concern for a longer GNSS outage that makes for inaccurate precisionpositioning. In the case of a longer outage, the application of machine-learning algorithms can help to maintain positioning accuracy [28].The Kalman filter (KF) is widely used in navigation as a data-fusion algorithm. In the integration of GNSS/INS, when a GNSS cannotnormally supply measurement updates, the filter time will be increased. In such cases, the divergence of strap-down inertial navigation system(SINS) or INS error occurs quickly without GNSS information correction. This could be detrimental depending on the use case, such as inthe concealment of unmanned underwater vehicles (UUVs) and unmanned aerial vehicles (UAVs). In [46], a backpropagation neural network(BPNN)-aided integrated navigation method based on vehicle motion learning was proposed. In [47], a recurrent neural network (RNN) incomparison with an extreme learning machine (ELM) and extended Kalman filter (EKF) was employed, while [48] used an NN model todesign a neural network-assisted GNSS/SINS calibration system. A CNN-based adaptive Kalman filter was designed and discussed in [49].In [50], ML is integrated with 3DMA (3D modelled assisted GNSS) to achieve high accuracy in urban environments. This is done usingML/AI algorithm together with Google’s vast database of 3D building models, the asymmetric NLOS propagations are modelled in order tocorrect NLOS pseudorange errors. This was shown to reduce Wrong-Side-of-Street occurrences from GNSS in phones up to 50 to 90%.In general, when a GNSS is active, the ML model is used to learn the divergence characteristics of the INS error under several basicconditions, depending on the area of application (vehicles, UAV etc.). If there is a disturbed GNSS signal, the ML model is used to correctthe position error of the INS in order to improve navigation accuracy. I. Satellite Selection Location accuracy is the result of two main factors, namely, the satellite location-dependent geometric dilution of precision (GDOP) andthe pseudorange measurement inaccuracies [51]. The benefit of multi-constellation GNSS is that there are more visible satellites to improveuser positioning performance. However, due to some issues, such as low-cost receivers having limited tracking receiver channels (four or fivechannels) and power consumption (also applicable to sophisticated receivers having up to 12 channels), it is usually not possible, beneficialor desirable to use all satellites in view for positioning [52]. Instead, an optimal subset is generally selected from all the possible visiblesatellite combinations with the aim of minimizing either the GDOP or weighted GDOP (WGDOP), as in the case of [53]. The GDOP iscomputed from the trace of the inverse of the measurement matrix of the contributing satellites [54]. Other selection criteria used in studiesinclude elevation angle, C/N0 and range errors. In [53], an end-to-end DL network for satellite selection based on the PointNet [55] andVoxelNet [56] networks were proposed. The satellite selection procedure is converted to a satellite segmentation problem that has two classlabels, one for selected satellites and the other for satellites not selected. In [54], the functional relationships between the entries of ameasurement matrix and the eigenvalues of its inverse are learned using a neural network, and the model is able to give the GDOP valueswithout the need to compute the inverse of a matrix. IV. WHAT ARE OTHER PROSPECTIVE GNSS AREAS IN WHICH ML ALGORITHMS COULD BE UTILIZED? There are other GNSS areas in which ML can be utilized, including space applications, such as robotic exploration in space and situationawareness for space debris analysis. ML models could be used in fault detection, altitude determination of space vehicles, flight control etc.Also, ML algorithms could be proposed to estimate products similar to international GNSS service (IGS) products using the IGS observationdata model to correct GNSS position error. Furthermore, ML algorithms could be proposed to emulate some of the output of a satellite-basedaugmentation system (SBAS) processing facility and predict the performance of SBASs in the context of a single frequency, singleconstellation (SFSC) and/or dual-frequency multi-constellation (DFMC) operation. These are areas that have seen much interest, especiallyin the latest European Space Agency (ESA) invitations to tender. They could also be used to detect signal diffraction, which has seen lessresearch. This effect occurs when GNSS signals are partially obstructed but not totally blocked. Signal diffraction is usually statisticallymodelled into the navigation algorithms, in which they are mitigated by smoothing techniques that adapt poorly. This makes these solutionsineffective, especially when there is an event too complex to be modelled with classical statistics and does not fit the mathematical assumptionsused to develop the statistical model.There are well-known models to estimate the precise position of a GNSS receiver (e.g., tropospheric and ionospheric models) undersimple scenarios with good line of sight, but there are high non-linearities if the receiver experiences multipath effects (e.g., urban scenarios,which have seen a high amount of research) or high signal dynamics (e.g., a rocket launch—in the domain of space applications). It would beinteresting to use deep learning in the later problem. ML could also be applied to space applications (space debris analysis, robotic explorationin space, cube satellites etc.). We have seen from this review that artificial neural networks assisting Kalman filter positioning in order todetermine the measurement noise covariance and process noise covariance of the Kalman filter have been promising and can be applied inother scenarios. V. ARE THERE ML ALGORITHMS THAT OUTPERFORM THE OTHERS? From the review done, ML algorithms were implemented in several GNSS use cases. It was observed that a GBDT-based algorithmachieved a classification accuracy higher than distance weighted k-nearest neighbour (KNN), traditional DT and an adaptive network-basedfuzzy inference system (ANFIS) [12] for GPS signal reception classification (LOS/NLOS/multipath). In [13], the proposed DT-basedclassifier obtained a higher accuracy than the KNN and SVM classification techniques. The DL models have also been seen to perform muchbetter; for example, [8], [9] and [11] showed good performance using CNN models. The next ML model with significant performance wasan NB prediction model, as seen in [7], in which the model performed better than logistic regression, SVM and traditional DT, according totheir studies. Another ML model with significant performance was the SVM in various forms (SVM, LSSVM-KF, C-SVM), as seen in [27],[28] and [38], respectively. As illustrated in Fig. 2, it was also noticed that the algorithms utilized in the GNSS use cases examined with better performance mostlyincluded SVMs, 11%; DTs (GBDT), 12%; and NN (or DL) models, 47% [30], [31], [32], [34], [33], and [35]. The distribution of the DLmodels used based on the literature reviewed is illustrated in Fig. 3. Fig. 2. A summary of the ML algorithms mostlyutilized in GNSS. From Fig. 3, we can see that CNN was implemented five times, followed by RNN and ANN with frequencies of three and two,respectively. VI. CHALLENGES AND RISKS OF USING ML ALGORITHMS IN GNSSS A primary source of risk is the data used to train the ML model. It is therefore important to ensure the reliability, integrity,representativeness and security of the training data being used. Otherwise, it is possible to produce false predictions and a bad ML model.Furthermore, “false correlation” is a risk that can occur when features are completely independent of each other and exhibit very similarbehaviour by chance. This can provide a false conclusion that they are somehow connected. An example is when a high C/N0 value isincorrectly attributed to multipath propagation, even though high C/N0 values can also occur during normal LOS situations. ML algorithmsare dependent on the data used for training and learning. Therefore, it is especially important to ensure the privacy and confidentiality of thedata used for building or training the ML models. Otherwise, data extraction attacks could be launched by hackers that jeopardize the entireML system. Since most ML systems take advantage of an already trained ML model, a generic model can be tweaked to meet a specific needby conducting specialized training. Therefore, what is referred to as a transfer learning attack can affect the task-specific ML model derivedfrom the generic model. Other risks include a non-representative sample used for training data, developers’ bias, system parameters, feedbackloops, overfitting/underfitting, misinterpretation of the resultant models and contaminated reference data or data bias.Lastly, because ML/DL algorithms may be hard to understand and explain (a “black box” effect, that is, a lack of transparency in somealgorithms), for example, how the weighted coefficients within a neural network are arranged to arrive at the correct answers, this may leadto difficulty in validating and explaining results to stakeholders in order to allow such models to be applied in real-life scenarios. VII. CONCLUSIONS AND FUTURE WORK Machine learning has been an enabling technology for many industrial applications in areas such as image recognition, classification,early anomaly detection and prediction. This review has shown that the utilization of ML in GNSSs has been an area of increasing researchinterest with positive prospects for future research and application.The most common use cases for ML in GNSS, according to the recent research, include signal acquisition, NLOS/multipath/evil waveform(EWF) detection, signal reception classification (LOS/NLOS/multipath signal classifier), earth observation/monitoring, GNSS positioningerror estimation, indoor navigation, characterization of GNSS ionospheric amplitude scintillation, detection of GNSS ionosphericscintillations/forecasting ionospheric time delay, GNSS phase scintillations, estimation of GNSS single frequency ionospheric delay,prediction of tropospheric wet delay, detection of GNSS spoofing attacks and jammer classification, GNSS/INS integration, and satelliteselection etc.We showed that among all ML methods, DL models (NNs) were the techniques that had been widely used by researchers most lately andprovided the best results, followed by DTs and SVMs.The challenges and risks in using ML in GNSSs were discussed and include a broad, but admittedly, selective sample of training data,developers’ biases, system parameters, feedback loops, overfitting/underfitting, misinterpretation of the resultant models and contaminatedreference data or data bias. The security of the training data and models was also identified as a risk, as they could be potentially manipulatedby hackers.Finally, we proposed other prospective areas of GNSSs to which ML algorithms could be applied, such as (but not limited to) spaceapplications (space debris analysis, robotic exploration in space, cube satellites etc.), predictions of IGS products through the use of IGSdata, SBAS performance prediction, GNSS hardware error detection/compensation, live tracking and tracing for supply chain and logisticsand smart port operation. REFERENCES [1] E. Kaplan and C. Hegarty, Understanding GPS/GNSS: Principles and Applications, 3rd ed. Artech House, 2017. [2] R. Malhotra, “A systematic review of machine learning techniques for software fault prediction,” Applied Soft Computing, vol. 27, pp. 504–518, 2015, doi:https://doi.org/10.1016/j.asoc.2014.11.023. [3] R. Sun, L.-T. Hsu, D. Xue, G. Zhang, and W. Y. Ochieng, “GPS signal reception classification using adaptive neuro-fuzzy inference system,” Journal ofNavigation, vol. 72, no. 3, pp. 685–701, 2019, doi: 10.1017/S0373463318000899. [4] P. Borhani-Darian and P. Closas, “Deep neural network approach to GNSS signal acquisition,” 2020, doi: 10.1109/PLANS46316.2020.9110205. Fig. 3. Distribution of the use of deep learning models inthe GNSS studies examined. [5] D. Borio, “Global positioning system: Signals, measurements and performance, second edition,” Ph.D. dissertation, 2006. [6] Y. Quan, “A new machine learning based method for multi-GNSS data quality assurance and multipath detection,” Ph.D. dissertation, University ofNottingham, 2017. [7] M. Socharoentum, H. A. Karimi, and Y. Deng, “A machine learning approach to detect non-line-of-sight GNSS signals in Nav2Nav,” 23rd ITS WorldCongress, 2016. [8] Y. Quan, L. Lau, G. W. Roberts, X. Meng, and C. Zhang, “Convolutional neural network based multipath detection method for static and kinematic GPS highprecision positioning,” Remote Sensing, vol. 10, 2018, doi: 10.3390/rs10122052. [9] E. Munin, A. Blais, and N. Couellan, “Convolutional neural network for multipath detection in GNSS receivers,” 2020, doi: 10.1109/AIDA-AT48540.2020.9049188. [10] L.-T. Hsu, “GNSS multipath detection using a machine learning approach,” 2017, doi: 10.1109/ITSC.2017.8317700. [11] A. Louis, “Neural network based evil waveforms detection,” 2019, doi: 10.23919/RFI48793.2019.9111769. [12] R. Sun, G. Wang, W. Zhang, L.-T. Hsu, and W. Y. Ochieng, “A gradient boosting decision tree-based GPS signal reception classification algorithm,” AppliedSoft Computing, vol. 86, 2020, doi: https://doi.org/10.1016/j.asoc.2019.105942. [13] B. Guermah, H. E. Ghazi, T. Sadiki, and H. Guermah, “A robust GNSS LOS/multipath signal classifier based on the fusion of information and machinelearning for intelligent transportation systems,” 2018, doi: 10.1109/ITMC.2018.8691272. [14] D. Brum et al., “A proposed earthquake warning system based on ionospheric anomalies derived from GNSS measurements and artificial neural networks,”in IGARSS 2019 - 2019 IEEE International Geoscience and Remote Sensing Symposium, 2019, pp. 9295–9298, doi: 10.1109/IGARSS.2019.8900197. [15] M. Alshaye, F. Alawwad, and I. Elshafiey, “Hurricane tracking using multi-GNSS-R and deep learning,” in 2020 3rd International Conference on ComputerApplications Information Security (ICCAIS), 2020, pp. 1–4, doi: 10.1109/ICCAIS48893.2020.9096717. [16] Q. Yan and W. Huang, “Sea ice sensing from GNSS-R data using convolutional neural networks,” IEEE Geoscience and Remote Sensing Letters, vol. 15,no. 10, pp. 1510–1514, 2018, doi: 10.1109/LGRS.2018.2852143. [17] Q. Yuan et al., “Deep learning in environmental remote sensing: Achievements and challenges,” Remote Sensing of Environment, vol. 241, p. 111716, 2020,doi: https://doi.org/10.1016/j.rse.2020.111716. [18] C. Ma, J. Yang, J. Chen, and Y. Tang, “Indoor and outdoor positioning system based on navigation signal simulator and pseudolites,” Advances in SpaceResearch, vol. 62, no. 9, pp. 2509–2517, 2018, doi: https://doi.org/10.1016/j.asr.2018.07.006. [19] P. Dabove and V. Di Pietra, “Towards high accuracy GNSS real-time positioning with smartphones,” Advances in Space Research, vol. 63, no. 1, pp. 94–102, 2019, doi: https://doi.org/10.1016/j.asr.2018.08.025. [20] A. Uzun, F. A. Ghani, H. Yenigün, and İ. Tekin, “A novel GNSS repeater architecture for indoor positioning systems in ISM band,” in 2020 IEEEInternational Symposium on Antennas and Propagation and North American Radio Science Meeting, 2020, pp. 1631–1632, doi:10.1109/IEEECONF35879.2020.9329653. [21] Y. Sun, J. Wang, and J. Chen, “Indoor precise point positioning with pseudolites using estimated time biases iPPP and iPPP-RTK,” GPS Solutions, vol. 25,no. 2, Jan. 2021, doi: 10.1007/s10291-020-01064-0. [22] W. Jiang, Y. Li, C. Rizos, B. Cai, and W. Shangguan, “Seamless indoor-outdoor navigation based on GNSS INS and terrestrial ranging techniques,” Journalof Navigation, vol. 70, no. 6, pp. 1183–1204, Jul. 2017, doi: 10.1017/s037346331700042x. [23] S. A. Platonov, A. V. Platonov, M. E. Postnikov, S. V. Khadonova, and S. S. Dymkova, “Using global navigation satellite systems to solve complexapplication problems,” in 2019 Systems of Signals Generating and Processing in the Field of on-Board Communications, 2019, doi: 10.1109/sosg.2019.8706807. [24] K.-W. Chiang, G.-J. Tsai, Y.-H. Li, Y. Li, and N. El-Sheimy, “Navigation engine design for automated driving using INS/GNSS/3D LiDAR-SLAM andintegrity assessment,” Remote Sensing, vol. 12, no. 10, p. 1564, May 2020, doi: 10.3390/rs12101564. [25] J. Li et al., “Real-time self-driving car navigation and obstacle avoidance using mobile 3D laser scanner and GNSS,” Multimedia Tools and Applications,vol. 76, no. 21, pp. 23017–23039, Nov. 2016, doi: 10.1007/s11042-016-4211-7. [26] Z. Fu, X. Feng, X. Duan, and Z. Fu, “An improved integrated navigation method based on RINS GNSS and kinematics for port heavy-duty AGV,”Proceedings of the Institution of Mechanical Engineers Part D: Journal of Automobile Engineering, vol. 234, no. 8, pp. 2135–2153, Feb. 2020, doi:10.1177/0954407019900031. [27] Z. Zhou, Y. Li, C. Fu1, and C. Rizos, “Least-squares support vector machine-based Kalman filtering for GNSS navigation with dynamic model real-timecorrection,” IET Radar, Sonar & Navigation, 2016. [28] A. Kuratomi, “GNSS position error estimated by machine learning techniques with environmental information input,” 2019. [29] Y. Xia et al., “Recurrent neural network-based scenario recognition with multi-constellation GNSS measurements on a smartphone,” Measurement, vol. 153,p. 107420, 2020, doi: https://doi.org/10.1016/j.measurement.2019.107420. [30] Y. Liu, Y. J. Morton, and Y. Jiao, “Application of machine learning to the characterization of GPS L1 ionospheric amplitude scintillation,” in 2018 IEEE/IONPosition, Location and Navigation Symposium (PLANS), 2018, pp. 1159–1166, doi: 10.1109/PLANS.2018.8373500. [31] N. Linty, A. Farasin, A. Favenza, and F. Dovis, “Detection of GNSS ionospheric scintillations based on machine learning decision tree,” IEEE Transactionson Aerospace and Electronic Systems, vol. 55, no. 1, pp. 303–317, 2019, doi: 10.1109/TAES.2018.2850385. [32] L. M. I, D. V. Ratnam, S. Raman, and G. Sivavaraprasad, “Machine learning algorithm to forecast ionospheric time delays using global navigation satellitesystem observations,” Acta Astronautica, vol. 173, 2020, doi: 10.1016/j.actaastro.2020.04.048. [33] R. Orus Perez, “Using TensorFlow-based neural network to estimate GNSS single frequency ionospheric delay (IONONet)”, Advances in Space Research,vol. 63, no. 5, pp. 1607–1618, 2019, doi: https://doi.org/10.1016/j.asr.2018.11.011. [34] K. Lamb et al., “Prediction of GNSS phase scintillations: A machine learning approach,” arXiv:1910.01570, 2019, doi: http://arxiv.org/abs/1910.01570. [35] M. O. Selbesoglu, “Prediction of tropospheric wet delay by an artificial neural network model based on meteorological and GNSS data,” Engineering Scienceand Technology, an International Journal, vol. 23, no. 5, pp. 967–972, 2020, doi: https://doi.org/10.1016/j.jestch.2019.11.006. [36] M. L. Psiaki and T. E. Humphreys, “GNSS spoofing and detection,” Proceedings of the IEEE, vol. 104, no. 6, pp. 1258–1270, 2016, doi:10.1109/JPROC.2016.2526658. [37] M. R. Manesh, J. Kenney, W. C. Hu, V. K. Devabhaktuni, and N. Kaabouch, “Detection of GPS spoofing attacks on unmanned aerial systems”, 2019, doi:10.1109/CCNC.2019.8651804. [38] S. Tohidi and M. R. Mosavi, “Effective detection of GNSS spoofing attack using a multi-layer perceptron neural network classifier trained by PSO,” in 202025th International Computer Conference, Computer Society of Iran (CSICC), 2020, pp. 1–5, doi: 10.1109/CSICC49403.2020.9050078. [39] R. Calvo-Palomino, A. Bhattacharya, G. Bovet, and D. Giustiniano, “Short: LSTM-based GNSS spoofing detection using low-cost spectrum sensors,” in2020 IEEE 21st International Symposium on A World of Wireless, Mobile and Multimedia Networks (WoWMoM), 2020, pp. 273–276, doi:10.1109/WoWMoM49955.2020.00055. [40] S. Semanjski, A. Muls, I. Semanjski, and W. De Wilde, “Use and validation of supervised machine learning approach for detection of GNSS signal spoofing,”in 2019 International Conference on Localization and GNSS (ICL-GNSS), 2019, pp. 1–6, doi: 10.1109/ICL-GNSS.2019.8752775. [41] S. Semanjski, I. Semanjski, W. D. Wilde, and S. Gautama, “Use of supervised machine learning for GNSS signal spoofing detection with validation on real-world meaconing and spoofing data —Part II”, Sensors, vol. 20, 2020, doi: 10.3390/s20071806. [42] R. M. Ferre, A. de la Fuente, and E. S. Lohan, “Jammer classification in GNSS bands via machine learning algorithms,” Sensors, vol. 19, 2019, doi:10.3390/s19224841. [43] F. Gallardo and A. P. Yuste, "SCER Spoofing Attacks on the Galileo Open Service and Machine Learning Techniques for End-User Protection," in IEEEAccess, vol. 8, pp. 85515-85532, 2020, doi: 10.1109/ACCESS.2020.2992119. [44] I. Belhajem, Y. B. Maissa, and A. Tamtaoui, “Improving vehicle localization in a smart city with low cost sensor networks and support vector machines,”Mobile Networks and Applications, vol. 23, no. 4, pp. 854–863, May 2017, doi: 10.1007/s11036-017-0879-9. [45] S. Joardar, T. A. Siddique, S. Alam, and M. Hossam-E-Haider, “Analyses of different types of errors for better precision in GNSS,” in 2016 3rd InternationalConference on Electrical Engineering and Information Communication Technology (ICEEICT), 2016, pp. 1–6, doi: 10.1109/CEEICT.2016.7873127. [46] Z. Zhang, F. Wu, Y. Liu, Y. Zuo, and Y. Ji, “A neural network aided integrated navigation algorithm based on vehicle motion mode information,” in 2019IEEE International Conference on Signal Processing, Communications and Computing (ICSPCC), 2019, pp. 1–5, doi: 10.1109/ICSPCC46631.2019.8960761. [47] H. Dai, H. Bian, R. Wang, and H. Ma, “An INS/GNSS integrated navigation in GNSS denied environment using recurrent neural network,” DefenceTechnology, vol. 16, 2020, doi: 10.1016/j.dt.2019.08.011. [48] Z. Liu, N. Wu, and F. Wang, “Application of neural network-assisted GNSS/SINS calibration system in UUV,” in 2019 IEEE 3rd Advanced InformationManagement, Communicates, Electronic and Automation Control Conference (IMCEC), 2019, pp. 1253–1257, doi: 10.1109/IMCEC46724.2019.8984176. [49] Z. Zou, T. Huang, L. Ye, and K. Song, “CNN based adaptive Kalman filter in high-dynamic condition for low-cost navigation system on highspeed UAV,”in 2020 5th Asia-Pacific Conference on Intelligent Robot Systems (ACIRS), 2020, pp. 103–108, doi: 10.1109/ACIRS49895.2020.9162601. [50] F. van Diggelen, Google's Use of 3D Building Models to Solve Urban GNSS, ITM/PTTI 2021 PLENARY SESSION. Available online:https://www.ion.org/itm/virtual-schedule.html. [51] T. Soininen, P. Syrjarinne, S. Ali-Loytty, and C. Schmid, “Data-driven approach to satellite selection in multi-constellation GNSS receivers,” 2018, doi:10.1109/ICL-GNSS.2018.8440912. [52] C. Park, I. Kim, J. G. Lee, and G.-I. Jee, “A satellite selection criterion incorporating the effect of elevation angle in GPS positioning,” Science Direct, vol.4, 1996, doi: https://doi.org/10.1016/S0967-0661(96)00192-X. [53] P. Huang, C. Rizos, and C. Roberts, “Satellite selection with an end-to-end deep learning network,” GPS Solution, vol. 22, 2018, doi: 10.1007/s10291-018-0776-0. [54] D. Simon and H. El-Sherief, “Navigation satellite selection using neural networks,” Neurocomputing, vol. 7, no. 3, pp. 247–258, Apr. 1995, doi:10.1016/0925-2312(94)00024-m. [55] C. R. Qi, H. Su, K. Mo, and L. J. Guibas, “PointNet: Deep learning on point sets for 3D classification and segmentation,” arXiv:1612.00593, 2017. [56] Y. Zhou and O. Tuzel, “VoxelNet: End-to-end learning for point cloud-based 3D object detection,” arXiv:1711.06396, 2017. I. Introduction II. Feasibility Assessment RQ1. To which GNSS use cases have ML algorithms been applied and in what are other prospective areas of GNSSs could ML algorithms be utilized? RQ2. Which ML algorithms are utilized in GNSSs and are there ML algorithms that significantly outperform the others? RQ3. What are the challenges and risks of using ML algorithms in GNSSs? III. Machine Learning Utilization in GNSS Use Cases A. GNSS Signal Acquisition B. Signal Detection and Classification C. Earth Observation and Monitoring D. GNSS Navigation and Precise Positioning E. Indoor Navigation F. Ionospheric Scintillations and Tropospheric Wet Delay G. GNSS Security—Spoofing and Jammer Attacks H. GNSS/Inertial Navigation Systems (GNSS/INS) Integration I. Satellite Selection IV. What are Other Prospective GNSS Areas in Which ML Algorithms Could Be Utilized? V. Are There ML Algorithms that Outperform the Others? VI. Challenges and Risks of Using ML Algorithms in GNSSs VII. Conclusions and Future Work References