1. Introduction
The positioning accuracy of a Global Navigation Satellite System (GNSS) is affected by a variety of systematic errors in its observations [
1], causing the processing of systematic errors to be a research hotspot. Although systematic GNSS errors have received considerable critical attention, no definitive method or model can eliminate these errors due to their complex spatiotemporal characteristics [
1]. Consequently, some of these errors remain unmodeled. These often-overlooked errors are called unmodeled errors [
2] and remain some of the most significant challenges for improving positioning accuracy.
Current research on the processing of unmodeled errors in relative positioning can be broadly divided into two categories [
3]. The first category studies specific types of errors, modeling them and correcting them more precisely to reduce the unmodeled errors remaining in the observation equations. Among such studies, the ionospheric delay, tropospheric delay, and multipath effect are the main objects of study. For the ionospheric delay, the processing method is selected according to the number of frequencies. Correction models can be adopted in single-frequency observations, for instance, the Klobuchar model [
4], NeQuick model [
5], and the BDGIM [
6]. In the case of multi-frequency observations, the effects of ionospheric delay can be significantly mitigated via a linear combination of observations [
7]. For tropospheric delay, its effect in each elevation angle direction is usually expressed as the product of the zenith delay and the mapping function. In terms of the zenith delay, models can be basically divided into two groups, geodetic-oriented models or navigation-oriented models, according to application. The first group is the most accurate but requires surface meteorological data, such as the Hopfield [
8], Saastamoinen [
9], and GPT2 [
10] models. The second group of models do not need surface meteorological data but are less accurate, such as the TropGrid2 [
11] and ITG [
12] models. Regarding the mapping functions, Marini [
13] derived an expression based on a continued fraction, and subsequently, many mapping functions have a similar form but different parameter values [
14]. As for the multipath effect, it can be mitigated by placing the receiver away from reflecting objects [
15] or by installing a choke ring for the antenna [
16]. In addition, the multipath effect can also be mitigated from the perspective of data processing, such as the sidereal filtering method [
17] and the hemispherical model [
18]. However, despite many studies focused on correcting a single error, unmodeled errors, such as the higher-order component of the ionospheric delay [
19] and the wet delay component of the tropospheric delay [
20,
21], are still not negligible, thus limiting the accuracy of positioning.
The second category of processing unmodeled errors in relative positioning is to study all systematic errors remaining in an observation equation, that is, to address the unmodeled errors as a whole [
3]. In that sense, a common approach is to treat them by including them in either a stochastic model or a functional model. Li et al. [
22] proposed a procedure to test the significance of unmodeled errors and suggested means of processing each type of unmodeled error. If the unmodeled errors are insignificant, they are included in the stochastic model. Zhang et al. [
23] proposed a method to dynamically determine the stochastic model using the satellite elevation angle and the carrier-to-noise power density ratio. Yuan et al. [
24] considered the influence of unmodeled errors and conducted a systematic study on stochastic models of low-cost receivers and smartphones. For standalone receivers, Zhang et al. [
25] proposed an unmodeled-error-corrected stochastic assessment regardless of the number of tracked frequencies. In contrast, when the unmodeled errors are significant, they need to be included in the functional model. Zhang and Li [
20] proposed a method based on multi-epoch partial parameterization to mitigate the unmodeled errors and verified its performance using six baselines from 4.82 km to 24.22 km. This method estimates the unmodeled error inside a moving window as a constant such that the variation in the unmodeled error between moving windows is ignored.
For time-varying unmodeled errors, developing an accurate model would be extremely beneficial so it may be incorporated into the positioning process for the real-time estimation of a much larger number of unmodeled error parameters. However, until now, little attention has been paid to unmodeled errors considering their time correlations [
1]. Therefore, the purpose of this paper is to investigate the problem of selecting basic functions for fitting unmodeled errors, which is a necessary step for providing a theoretical basis and practical suggestions for the further development of specific time-varying models. To achieve this goal, we compare three basic functions, polynomials, sinusoidal functions, and combinatorial functions, to fit the unmodeled errors in differential positionings. The experiment data were obtained from the International GNSS Service (IGS) and the US National Oceanic and Atmospheric Administration (NOAA) UFCORS Center, and comparisons of different basic functions are carried out in terms of residuals, overall accuracy, and processing time.
The rest of this paper is organized as follows. The source and preprocessing of unmodeled error data are first introduced. Then, three alternative basic functions and corresponding methods are identified. Next, the results of fitting experiments and positioning experiments are analyzed and discussed. Finally, the conclusions are summarized.
4. Discussion
Previous studies [
3,
23] indicated that the unmodeled errors contain periodic components. The frequency range of these periodic components is mainly
Hz, corresponding to a time range of about 200~2500 s. In this paper, when fitting the unmodeled errors with the sinusoidal or combinatorial functions, the chosen frequency ranges were also concentrated in the
Hz range. However, from
Figure 4, it can be found that a good result cannot be obtained when a sinusoidal function alone is used to fit the unmodeled errors. Meanwhile, from
Figure 6 and
Figure 7 and
Table 5, it can be found that the addition of sinusoidal functions improves the fitting effect but quite minimally when using a combinatorial function. From these experimental results, we can conclude that the periodic components of the unmodeled errors represent a minor proportion and that most of the components behave as trend term signals.
Previous studies also concluded that atmospheric delays are the main component of GNSS unmodeled errors. Meanwhile, noting that the unmodeled error time series used in this paper lasted 3 h, the experiments also imply that the atmospheric delays are relatively stable within 3 h. The periodic components ( Hz) may be mainly attributed to the multipath effect and some uncorrected receiver bias and are not the main proportion of unmodeled errors. This explains why the unmodeled errors can be fitted well using polynomial functions alone and limits the conclusion of this paper to the short-term modeling of unmodeled errors. This application is relevant to real-time navigation.
Another reason for the minor proportion of the periodic components in the unmodeled error may be related to the estimation of the unmodeled errors. The main source of the unmodeled errors obtained from the inversion in this paper was the component in the positioning and the ambiguity parameters. During the solving process, it is possible that the errors absorbed by the static positioning and the ambiguity parameter were mainly non-periodic components. Most of the periodic components were absorbed by other parameters. These parameters cannot be considered in the inversion method of this paper because it is difficult to obtain their true values.
In the positioning experiments, there is a significant improvement in the positioning results corrected by the fitted unmodeled errors. This indicates that the fitted, unmodeled errors were consistent with the actual unmodeled errors in the observation equations. There was still some deviation between the improved positioning results and the true values because the original modeled errors were still present in the observations. When positioning with observations corrected for unmodeled errors, these originally modeled errors affected the parameter estimates in the current epoch. However, these modeled errors were smaller compared to the errors in the uncorrected equations, explaining the larger improvement over the original positioning results, although the corrected results still deviated somewhat from the true values.
5. Conclusions
The present study addressed some basic functions to account for unmodeled errors in medium and long GNNSS baselines in two differential positioning modes: estimating atmospheric delays and using the IF combination. We used four baselines ranging from 30 to 220 km to invert an unmodeled error time series to provide a reference residual. We then selected three alternative basic functions including polynomials, sinusoidal functions, and combinatorial functions, and methods for their solution, testing, and evaluation methods were introduced. At last, fitting experiments and positioning experiments were conducted to analyze and compare these three basic functions.
The experimental results can be summarized as follows. (1) When fitting unmodeled errors with polynomials in a long baseline, the second-order polynomial is the optimal choice. Compared with the first-order polynomial, the second-order polynomial has higher accuracy and a faster convergence of residuals. Compared with higher-order polynomials, the second-order polynomial is more efficient in the case of ensuring comparable accuracy. (2) It is difficult to achieve high-accuracy results by fitting the unmodeled errors with sinusoidal functions only. (3) The combinatorial functions have basically no advantage compared with the second-order polynomial in short-term cases. When estimating atmospheric delays or using the IF combination, the combinatorial functions are essentially comparable to the second-order polynomial in terms of residual convergence time and overall accuracy but takes several times longer than the second-order polynomial. (4) After the observations have been corrected by the fitted unmodeled errors from the second-order polynomial, a remarkably significant improvement appears in terms of the positioning results, which demonstrates the usability of fitting the unmodeled errors for positioning using the second-order polynomial.
Based on the above experimental findings, we suggest that a second-order polynomial can be used as the basic function for a short-term time-varying model of unmodeled errors in medium and long baselines. Further work needs to be carried out in order to add the basic functions into the positioning process of long baselines to estimate model parameters in real time.