Empirical analysis and modeling of Argos Doppler location errors in Romania

Introduction

Advances in wildlife tracking technologies allow researchers to track the movement of many terrestrial and aquatic species (Thomas, Holland & Minot, 2012). Movement analysis has evolved from short-term local studies on small numbers of individuals to long-term global studies on hundreds of individuals, allowing researchers to answer more complex questions about animal movement and space use (Block et al., 2011; Sequeira et al., 2018). These data can be included in statistical models and used to understand movement patterns, population redistribution, habitat use, habitat selection, and conservation needs (Bridge et al., 2011; Doherty et al., 2017; Hooten et al., 2017; Pendoley et al., 2014; Pop et al., 2018; Schofield et al., 2013).

Collecting good quality movement data remains a challenging task mainly due to technological constraints (Hooten et al., 2017). Well-known tracking technologies such as radio telemetry (VHF telemetry), satellite-based telemetry (GPS, Argos), and light-level geolocation have certain limitations (Bridge et al., 2011). The main challenges are the physical size and the mass of the devices. In particular, the mass of the device must not exceed 5% of the animal’s body-weight (Silvy, 2012). Furthermore, transmitters must be protected from environmental hazards and damage and must include a long-lasting battery or alternative power source for consistent one-way or two-way communication (Bridge et al., 2011). As such, devices meeting these parameters can be cumbersome and heavy (Silvy, 2012), and not well suited for many small-bodied animals.

The most accurate available tracking technology is the global positioning system (GPS), capable of a horizontal location accuracy of under 10 m (Madry, 2015). The weight of present-day GPS receivers varies between a minimum of four g (lifespan limited to a few transmission days and suitable for individuals weighing ∼80 g) to in excess of one kg (typically a lifespan averaging 2 years and suitable only for large animals). An alternate option for long-term animal movement studies is the Argos satellite Doppler-based system, which relies on transmitters with extended lifespans that now weigh less than four g, and are able to deliver an unlimited number of locations in near real-time (Bridge et al., 2011; Hooten et al., 2017; Thomas, Holland & Minot, 2012). However, the small size comes at a cost in terms of accuracy of localization compared to GPS. Thus, data interpretation may pose a challenge for inexperienced users (Rozylowicz et al., 2018). Regardless of the device size, Argos transmitters, or Platform Transmitter Terminals (PTT), provide locations with the same error rate. In addition, the data requires the application of complex control processes, such as filtering and modeling (Thomas, Holland & Minot, 2012). If the PTT’s are equipped with GPS receivers, the location precision can be increased by retaining only validated locations (Lopez et al., 2015). Adding a GPS unit to a PTT device results in increasing the minimum device weight to approximately 20 g per unit.

The Argos system utilizes positioning instruments on board six polar-orbiting satellites, in addition to ground-based receiving stations and data processing centers. It delivers near real-time localization of transmitters to end-users. Argos locations are calculated from the Doppler shift of a PTT radio frequency during a satellite pass (CLS, 2016). Collecte Localisation Satellites (CLS), the operator of the Argos system, provides several metrics for data quality, such as a location class (LC) based on the number of messages received for each location. The estimated upper bound errors are 250 m for LC 3 (best accuracy class), 500 m for LC 2, 1,500 m for LC 1, and over 1,500 m for LC 0. For locations derived from three or fewer messages (LC A, LC B), Argos does not provide error thresholds. Invalid locations are labeled as LC Z and GPS locations as LC G (CLS, 2016). CLS pre-processes these locations using one of Argos’s nominal filters such as the Least squares algorithm or the Kalman filter (Lopez et al., 2015). In practice, location errors of 10–100 km often occur due to communication conditions driven by the environment or animal behavior (e.g., animal speed, terrain fragmentation, rain, cloud cover, temperature) (Christin, St-Laurent & Berteaux, 2015; Costa et al., 2010; Douglas et al., 2012; Dubinin, Lushchekina & Radeloff, 2010; Sauder, Rachlow & Wiest, 2012; Witt et al., 2010). Thus, filtering the data to exclude implausible Argos locations before employing movement analysis has become a standard approach for researchers (Hooten et al., 2017). Furthermore, the quality of data seems to be highly dependent on geographic location. The Argos transmission systems in Eastern Europe are lower in power, their signals being hidden by a background radio noise present across the Argos frequency (Gros, Malardé & Woodward, 2006).

Location errors can be filtered using destructive (i.e., removing implausible locations) and reconstructive filters (i.e., evaluation of uncertainty in the estimation of locations) (Douglas et al., 2012). Destructive filters remove duplicates (e.g., identical timestamps), locations outside of a defined range (e.g., thresholds for geometric dilution of precision, latitude, longitude, LC), or locations exceeding a fixed movement rate or a turning angle (Douglas et al., 2012; Kranstauber et al., 2011). One such advanced destructive filter is the Douglas Argos filter algorithm, available on the Movebank database of animal tracking data (Kranstauber et al., 2011). The Douglas Argos filter algorithm uses thresholds to mark the outliers as implausible locations. It is available in three settings: the maximum redundant distance filter (MRD) (which retains near-consecutive locations within a distance threshold), the Douglas Argos distance angle filter (DAR) (which retains near-consecutive locations within a distance threshold and location-passing movement-rate and turning-angle tests), and the hybrid filter (HYB) (which combines MRD and DAR parameters specifically for migratory species) (Douglas et al., 2012). In contrast, reconstructive filters employ advanced statistical methods to detect animal movement characteristics without removing locations (e.g., discrete‐time movement model, correlated random walk state-space models, movement-based kernel density estimates, Bayesian State-Space Models, and hidden Markov models), or model the data using the errors associated with movement (e.g., Argos error ellipse) (Hooten et al., 2017; Jonsen, Flemming & Myers, 2005; Lopez et al., 2015; Silva et al., 2014).

Considering the behavioral, environmental, and geographic specificity of the errors associated with Argos data and the lack of benchmark studies in Eastern Europe, the research objectives for this study are: (1) to provide empirical evidence of the accuracy of locations collected via the Argos Doppler system in Romania, (2) to investigate the effectiveness of straightforward destructive filters for improving Argos data quality, and (3) to provides guidelines for processing Argos wildlife monitoring data in Eastern Europe. The errors associated with Argos locations were assessed in four geographic locations from Romania under three spatial movement conditions: static, low-speed and high-speed. The effectiveness of the Douglas Argos distance angle filter algorithm was then evaluated in terms of location error minimization.

Materials and Methods

Trial sites

The accuracy of Argos locations was analyzed in four areas in Romania. These areas varied both topographically and in terms of reception conditions and could be categorized as:

1. Urban: two urban parks within a residential area of Romania’s largest metropolis, Bucharest (within the Tineretului Park for the static test, 44°24′N 26°06′E; and the shoreline of Vacaresti Lake for the mobile tests, 44°24′N, 26°07′E).

2. Unobstructed flat rural lowland: Saveni, Ialomita County (44°35′, 27°37′E).

3. Highly fragmented rural topography: Iron Gates Natural Park, Mehedinti County (44°41′N 22°21′E), along the Danube River.

4. Moderately fragmented rural upland: Sighisoara (within Breite for the static tests, 46°12′N 24°45′E; and Sighisoara-Apold for the mobile tests, 46°09′N 24°46′E)).

At all trial sites, three tests were carried out: a static test, a low-speed test, and a high-speed test (Fig. 1).

Results

Between June 2017 and September 2017, the five PTTs received 3,705 valid Argos locations (Data S1). Each PTT generated a similar number of locations (min = 717, max = 760, χ² (df = 4, n = 3,705) = 1.86, p = 0.76). For each location, the Argos PTTs transmitted between 1 and 14 messages to one of the six polar-orbiting satellites fitted with Argos instruments. The Argos satellites resulted in the calculation of a dissimilar number of locations (min = 310, NOAA-N, NP′; max = 890, NOAA-18, NN, χ² (df = 5, n = 3,705) = 399.87, p < 0.001).

The dataset was dominated by low-quality data, with over 29% of locations labeled as LC B. 46% of the locations were classified by the CLS as error bounded (Argos LC 3, 2, and 1), from which 14.25% were of high estimated quality (LC 3, <250 m estimated accuracy) (Table 1).

The empirical mean location error for the five PTTs was 3,583.66 m (stdev = 8,225.96 m). Location errors differed significantly by Argos LCs (Kruskal–Wallis chi-squared = 1,170.95, df = 5, p < 0.001), except for the LC 1 and LC A which showed identical ranking. All the error-bounded LCs (Argos LC 3, 2, and 1) had measured errors which were significantly larger than the location-class-specific 68th percentile estimated by CLS. Argos LC A and LC B had smaller associated errors than LC 0, a LC which is considered more accurate by CLS (Fig. 2; Table 1).

On average, longitudinal errors were larger (mean = 2,872 m, stdev = 7,678 m) than latitudinal errors (mean = 1,604.4 m, stdev = 3,272.3 m). This pattern was found to be consistent in all LCs (Table 1). The largest proportion of longitudinal errors were in LC 0 (73.96% of locations) and LC 1 (73.44% of locations), while the largest proportion of latitudinal errors was found in the LC B (39.29%) and LC A (38.42%) classes. Geographically, most errors were to the East and West, followed by the North-East and South-West. Those oriented toward the North and South were less present in the dataset (Figs. S1 and S2).

The best linear mixed model showed that the fixed factors Motion (speed tests) and Place (trial areas) had a significant impact on the errors (Table S1). The fixed factors Motion and Place accounted for 17.45% of the variance in the error data (marginal R-squared), while the fixed and random structures (reception points nested in the satellite) combined accounted for 56.79% of the variance (conditional R-squared). The proportion of variance in errors accounted for by the reception points nested in the satellite (Intra Class Correlation Coefficient) was 41.91%, while the satellites themselves accounted for only 5.73%, suggesting that reception conditions at transmission time strongly influenced the quality of the data (Table 2).

A comparison of the confidence intervals of the fixed factors showed that the locations from motion-controlled trials differed significantly. Errors from high-speed tests were larger than those in the low-speed tests, and errors in the low-speed tests were larger than in the static tests. The mean error for the static tests was 2,708.84 m, the mean error for the low-speed tests was 3,779.73 m, and the mean error for high-speed tests was 4,550 m (Fig. 3; Table 3). The trial sites also contributed to the error variance, with locations from Iron Gates (a narrow valley) generating larger errors when compared to locations from the other sampling areas. As a comparative example, the mean error for Iron Gates was 4,698.35 m and the mean error for Saveni was 3,122.01 m (Fig. 3; Table 4). Interestingly, it was found that substantial errors existed in the static tests from Iron Gates which were as large as the location errors associated with the low-speed tests conducted in Saveni, Bucharest, and Sighisoara (Fig. 3).

The DAR filter applied to raw data, successfully excluded the largest errors when MAXREDUN was defined for local (two km, DAR 2) and regional (15 km, DAR 15) scales of study. However, the DAR 2 filter was more effective at excluding large errors, by retaining 84.35% of the initial locations compared to 94.82% for the DAR 15 filter (Tables S2 and S3). The mean error for the DAR 2 filtered data was 2,313.51 m (stdev = 3,134.67), a 35.45% improvement compared to the error associated with the raw data, while the improvement after the DAR 15 filter was applied was only 27.05%. The DAR 15 filter retained almost all the locations in LC 3, LC 2, and LC 1 classes, while the DAR 2 filter altered the number of LC 2 and LC 1 locations slightly. The most heavily-impacted LC was LC 0—the class with the most substantial errors amongst the data—with only 68.35% of locations retained by the DAR 2 filter and 90.42% by the DAR 15 filter (Tables S2 and S3; Fig. 4). The intermediate thresholds (DAR 5 and DAR 10) did not improved the data accuracy in comparison to DAR 2, and the results were only marginally different when compared to DAR 15 (Fig. S3). Douglas Argos algorithms filtered longitudinal and latitudinal errors equally; thus, the filtered data are also anisotropic, with the longitudinal errors larger than the latitudinal errors (Fig. 5).

Discussion

The accuracy of the Argos Doppler locations received from Romania was negatively influenced by the movement speed and topographic characteristics of the trial sites. The empirical data showed that Argos locations yielded a lower accuracy, even in stationary tests performed in unobstructed areas. This suggests that Argos Doppler telemetry data must undergo a comprehensive filtering process before being used in movement analysis.

In our experiment, 14% of locations were considered category LC 3, the most accurate Argos LC (CLS, 2016). However, the 68th percentile of LC 3 locations was twice as large as the 68th percentile provided by CLS as the upper bound error (520.85 vs. 250 m). All error-bounded Argos locations classes (LC 2, LC 1, and LC 0) included larger positional errors than those indicated by CLS (2016), which is in line with the results demonstrated in other controlled and real-life studies. For example, a stationary and mobile test in Southern Russia (Dubinin, Lushchekina & Radeloff, 2010) and tests on marine species (Costa et al., 2010) yielded errors similar to ours for LC 3 data. Data recovered from these other studies and our results indicate higher errors than those indicated by CLS for all LCs. LC 0 was the most inaccurate LC (68th percentile = 5,877.38 m), which corroborates the results of other studies (Douglas et al., 2012; Lowther et al., 2015). This suggests that LC 0 locations must be filtered together with LC A and LC B, and should not be considered an accurate LC. Argos errors are not isotropic, and longitudinal errors were larger than latitudinal errors (CLS, 2016; Douglas et al., 2012), as already reported by all benchmark studies (Lowther et al., 2015; Sauder, Rachlow & Wiest, 2012; Witt et al., 2010). In this study, the mean latitudinal errors for LC 3, LC 2, and LC 1 were only slightly larger than the CLS 68th percentile for the respective LCs. However, these data are not likely to be useful for movement studies, such as home-range analysis (Hooten et al., 2017) since longitudinal errors were significant even in optimal reception environments (e.g., flat areas, unobstructed by vegetation). South-Eastern Europe is considered an area with poor reception quality due to the broadband noise covering the Argos 401.65 MHz ± 30 kHz frequency (Gros, Malardé & Woodward, 2006), which might have a negative impact on quantity and quality of data. Since these errors were similar to those obtained in other studies outside of Europe, the broadband noise affecting Southeastern Europe (Gros, Malardé & Woodward, 2006) seems to have had only a minimal influence on the accuracy of Argos data.

The accuracy of Argos Doppler locations is influenced by a plethora of factors such as the PTTs’ repetition rate, topography, vegetation, terrain ruggedness, electromagnetic noise, and geographic area (Freitas et al., 2008; Lowther et al., 2015; Nicholls, Robertson & Murray, 2007; Sauder, Rachlow & Wiest, 2012). The best linear mixed effects model for this study showed that the movement speed of the PTT had the most significant influence on Argos location errors, while the trial area contributed only marginally to the variation in errors. As expected, motion-controlled tests generated significantly varying errors. Static tests generated smaller errors than low-speed tests, and high-speed tests generated larger errors than low-speed tests. However, topographical obstruction of the sky influenced data acquisition. Data obtained from the highly-fragmented Iron Gates trial area contained larger errors than the other three trial areas, including Bucharest city potentially affected by electromagnetic interference (Gros, Malardé & Woodward, 2006). In the Iron Gates area, the static test generated positional errors as large as in the low-speed motion tests performed in the other three areas. This suggests that locations from fragmented areas may be highly imprecise and can lead to biased conclusions about animal movement and location if not adequately filtered (Lopez et al., 2015). The variance explained by the random part of the linear mixed effects model suggests that the satellite detecting the location of the PTT had a minimal impact on accuracy, while the dominant source of positional errors were probably due to other random factors, such as poor line of sight to the satellite as a result of local topography, the presence of obstructing vegetation or the relative orientation of the respective PTT with respect to the sky (Christin, St-Laurent & Berteaux, 2015; Doherty et al., 2017; Dubinin, Lushchekina & Radeloff, 2010; Soutullo et al., 2007).

Due to the large positional errors, Argos Doppler data should be filtered or modeled, considering the uncertainty of locations (McClintock et al., 2014). Data filtering is a challenge, as the aim is to reduce low-quality data as much as possible while retaining the necessary amount of data for analysis (Hooten et al., 2017). In the filtering exercise conducted in this study, we tested the effect of the Douglas Argos distance, angle, and rate filter. This filter retains spatially redundant locations, passing movement rates, and turning angle tests (Douglas et al., 2012). The results indicated that the selection of a proper self-validating distance threshold significantly reduces the errors while retaining a large amount of data. In this study, a larger threshold, MAXREDUN = 15 km, reduced the efficacy of the filter considerably by retaining 10% more locations than when the threshold was set at two km. The differences between the two approaches suggested that previous knowledge of movement behavior of species of interest is essential to ensure good quality data. For example, if the species of interest is known to perform frequent long-distance movements, then a larger MAXREDUN is required. The DAR filter was tested by targeting all the LCs; however, LC 3, LC 2, and LC 1 were only slightly impacted, and thus it is recommended to run the filter using LC 1 as the threshold LC as suggested by Douglas et al. (2012).

Even if selecting the optimal threshold, the post-processed data may include large positional errors; therefore, we recommend incorporating Argos error metrics such as error ellipse into models (McClintock et al., 2014). As an alternative, complex statistical approaches such as state-space modeling can be used instead of standard movement analysis (Hooten et al., 2017). While the results were provided based on the distance, angle, and rate filter which fitted the data, other available filtering approaches might be more effective for a given species (e.g., speed filters, Douglas Argos MRD, Douglas Argos HYB).

Conclusions

To provide guidance for processing Argos Doppler-derived locations obtained from Eastern Europe, we assessed the errors associated with Argos locations in four geographic locations in Romania using static, low-speed and high-speed tests. The effectiveness of destructive filters was evaluated in terms of the minimization of location errors, using the Douglas Argos distance angle filter algorithm as an example. The results indicated that the received Argos locations had larger positional errors than those indicated by the operator of the Argos system. This included cases where the reception conditions were ideal. The magnitude of the errors varied among LCs; however, locations in the LC 0 class were prone to large errors. Positional errors were anisotropic, predominantly oriented East and West, which resulted in larger longitudinal errors. Errors were mostly related to the speed of the transmitters and the reception conditions at the time of transmission (e.g., the orientation of the transmitter with respect to the sky), but other factors such as topography contributed to receiving low-accuracy data as well. The destructive Douglas Argos distance angle filter removed between 15% (self-confirming distance threshold = two km) and 5% (threshold = 15 km) of locations. However, the mean errors remained larger than those indicated by the operator of Argos system. The Argos data should therefore be used with caution in movement ecology studies, especially for species with small home ranges, such as songbirds, reptiles, or small mammals. Filter selection for data processing requires knowledge about the movement patterns and behaviors of the species of interest, and parametrization of the selected filter must follow a trial and error approach. Argos Doppler is unsuitable for movement analysis of species with limited movement capabilities. GPS tags would be preferable for such species, enabling detailed analysis of habitat selection and movement, with the caveat that store-on-board systems might be required owing to the larger sizes of the units that send data remotely via satellite.

Supplemental Information