Validating a high-throughput tracking system: ATLAS as a regional-scale alternative to GPS

Fine-scale tracking of animal movement is important to understand the proximate mechanisms of animal behaviour. While GPS tracking is an excellent tool for measuring animal movement, trade-offs between tag weight, cost and lifespan limit its application to relatively large species, a small number of individuals or short tracking durations, respectively. The reverse-GPS system – ATLAS – uses lighter, cheaper tags compared to GPS tags, that can also last long periods of time at high sampling frequencies. Six systems are now operational worldwide and have successfully tracked over 50 species in various landscape types. This growing use of ATLAS to track animal movement motivates further refinement of best-practice application and an assessment of its accuracy. Here, we test the accuracy and precision of the largest ATLAS system, located in the Dutch Wadden Sea using concurrent GPS measurements as a reference. This large-scale ATLAS system consists of 26 receivers and covers 1326 km2 of intertidal region, with almost no physical obstacles for radio signals, providing a useful baseline for other systems. To measure accuracy, we calculated the distance between ATLAS and GPS location estimates for a route (mobile test) and 16 fixed locations (stationary test) on the Griend mudflat. ATLAS-derived location estimates differed on average 4.2 m from GPS-estimated stationary test sites and 5.7 m from GPS tracks taken whilst moving between them. Signals that were collected by more receiver stations were more accurate, although even 3-receiver localisations were comparable with GPS localisations (∼10 m difference). Higher receiver stations detected the tag at longer distances. Future ATLAS users should consider the height of receivers, their spatial arrangement, density and the movement mode of the study species (e.g., ground-dwelling or flying). In conclusion, ATLAS provides an accurate, regional-scale alternative to global GPS-based tracking with which hundreds of relatively small-bodied species can be tracked simultaneously for long periods of time. Our study shows that ATLAS is a valid alternative, providing comparable location estimates to GPS.


Introduction
Since the advent of tracking technology, insight into the often cryptic movements of animals has 48 helped us to understand behaviours that were previously almost impossible, from the migration 49 patterns of whales (Abrahms et al., 2019) and birds (Gill et al., 2009)  covariates (Eikelboom et al., 2020). One potential alternative for regional-scale studies is ATLAS 65 (Advanced Tracking and Localisation of Animals in real-life Systems; Toledo et al., 2020), a high-66 throughput system that uses low-cost and lightweight radio-transmitters to track animals at a regional 67 scale. 68 triangulation. Unlike GPS, radio tags act as transmitters rather than receivers, alleviating energy-73 demanding position computations and remote data communications. Miniature radio-transmitters have 74 low power requirements (thus requiring smaller batteries), increasing their potential for use with 75 smaller species while keeping cost minimal. However, conventional radio telemetry is labour 76 intensive and it is not feasible to follow more than a few individuals and locations are estimated 77 irregularly. Attempts to automate wildlife tracking based on standard radio tags were made as early as 78 the 1960's (Cochran et al., 1965) and more recently using a bounding array of receivers distributed 79 within a specific region (Kays et al., 2011). Locations can be estimated from the data collected by 80 these receivers and has the potential to be accessed almost instantaneously. However, while previous 81 system. ATLAS therefore provides high-throughput monitoring for relatively cheap (~€50 per tag), 90 lightweight (<1g + battery weight), and long-lasting (~8 months for 4.4g tag at 1/6 Hz) tags. 91 However, ATLAS installation requires time, resources and expertise, and its spatial coverage is 92 limited to a regional scale where the line-of-sight from three or more receiver stations overlap. tracking migrant shorebirds while they stopover in the area (Fig. 1). Eleven receivers were built on 130 temporary scaffolds on the mudflats and were powered with four 100 W solar panels and a 100 W 131 wind turbine (Ampair) connected to three (105 Ah) batteries. The remaining 15 receivers were 132 installed in places such as on buildings and other stable structures where power was available. Each 133 receiver had a UHF antenna connected to the radio through a custom-built front-end unit (CircuitHub) 134 and a custom built Low-Noise Amplifier. Radio-frequency samples from tags are processed by the 135 receiver's computer to estimate the arrival times of the signal. All receiver stations were connected to 136 internet using a 3G cellular model (USB dongles, Huawei E3372) to send detection reports to a 137 central server situated at the NIOZ. In real-time, the server calculated location estimates and stored 138 these in a MySQL (v5.7, https://www.mysql.com/) database. 139 140

Data collection 141
To test the accuracy of the WATLAS deployment and ATLAS tracking in general, we focussed our 142 tests on the area around Griend, where we study shorebird movement and have the highest 143 concentration of receivers (Fig. 1). We tested reception and localisation accuracy at 16 sites around 144 the mudflat, 9 of which were 1 km apart with an outer ring of 7 sites that were 2 km between each 145 other and the 1 km sites (Fig. 2). Between 21 st -27 th August 2020, we travelled to these sites while 146 carrying a handheld Garmin Dakota 10 GPS (<10 m error 95% typical) -set to record tracks on 'auto' 147 which records at a variable rate to create an optimum representation of tracks -and an ATLAS tag. 148 The tag we used for testing emitted a radio signal at 1 Hz. It consisted of a miniature frequency-shift-149 keying 434MHz integrated radio transceiver and microcontroller (Texas Instruments CC1310) and a 150 monopole ¼ λ gold-plated, multistranded steel wire antenna (Toledo et al., 2014). The tag was 151

158
We walked to 13 of the sites during the day at low tide on the 21 st , 23 rd and 27 th August (Fig. 2). 159 Between sites, the pole holding the tag was kept upright meaning that the tag was consistently ~1.2m 160 from the ground. On arrival at each of these sites, we pushed the pole into the sand so that the tag was 161 1m above the mudflat and attached the GPS to the pole. We collected ATLAS and GPS data at each 162 site for 5 minutes. Due to weather and time constraints, we were unable to travel to each site on foot. 163 On the 24 th August, we sailed a rubber boat to the some of the furthest sites in the western and 164 pole with the ATLAS tag and GPS tag attached was held upright on the pole in the middle of the boat 166 therefore the tag was ~1.3m above the water level. Despite being at anchor with a taut rope, there 167 were waves and therefore the boat was not completely stationary during the test, therefore we expect 168 slight overestimates in the location error and larger standard deviations for these positions. Easting and Northing (VARX and VARY). We removed localisations that had high VARX and 176 VARY (>2000) and then smoothed the data by computing a 3-point median smooth across the 177 localisations (Appendix S1). 178

Figure 2. Stationary test sites (numbered) on the Griend mudflat and routes between them 179
which are mapped using GPS data. Due to time and weather constraints, we used a boat to test 180 the western side of the mudflat (purple) but the rest was walked (green). 181

Stationary Test 186
For the stationary test, we calculated the mean GPS-derived location estimate at each of the 16 test 187 sites and compared it to the ATLAS-derived location estimate; hence, our measure of accuracy (which 188 we henceforth call 'error') refers to the difference in metres between the two estimated locations. For 189 each site, the mean error (m), standard deviation, the median error (m) and the 95 th upper and lower 190 percentiles. We investigated the tag reception at each location by calculating the fix rate (number of 191 localisations/300 (max possible localisations in 5-minute period at 1Hz) and the mean and standard 192 deviation of the number of receivers that contributed to each location estimate. 193

Mobile Test 194
When the tag was mobile, we compared each ATLAS-derived location estimate to the nearest (in 195 time) GPS-derived location estimate. We removed pairings of location estimates where the smallest 196 time difference between an ATLAS location estimate and a GPS location estimate was >2 seconds. 197 We calculated the distance (m) between the remaining paired locations to determine 'error'. As in the 198 stationary test, we calculated the mean error (m), standard deviation, the median error (m) and the 95 th 199 upper and lower percentiles to assess accuracy. We aggregated these summary statistics by the 200 number of receivers that contributed to each location estimate to assess the influence of the number 201 receiver stations on accuracy and investigate the overall coverage of our system. To assess tag 202 reception around the mudflat for specific receivers, we plotted each receiver station and the location 203 estimates that they contributed to on separate maps. 204 205

Results
For the stationary test, we calculated a median error [2.5-97.5%] of 4.2 m [0.6-110.1] over all sites, 208 which reduced to a median 3.1 m [0.5-27.5] after filter-smoothing the data (Fig. 3). When analysed 209 individually, we found that the least accurate sites were the three most Northern sites (1-3). These 210 sites were outside of the array of receivers and each had a mean of <4 receiver stations detecting them 211 (Fig. 3, Table 1). Fifteen out of sixteen sites had a fix rate of >90%. One site (site 2) had a fix rate of 212 73% and the lowest accuracy (median error = 110.5 m [5.5-1059.4]) and mean number of receivers 213 detecting the tag there (mean error ± s.d. =3.4 ± 0.8). This low accuracy was able to be mitigated by 214 applying the filter-smooth, which increased accuracy to 28.2 [7.6-67.5]. It should be noted however 215 that for this particular site the number of location estimates that remained after smoothing was only 216 9.7% of the expected number of points (300). 217  Table 1. Accuracy (error = distance (m) to GPS-derived location estimate) and fix rate (% localisations at 1Hz, max 300) of ATLAS tags (both raw 223 data and filter-smoothed data) that were stationary over five minutes. IDs relate to locations on Fig. 2. ID1, ID4 and ID12  and accuracy increased with receiver number (Table 2). Larger errors were more likely to occur on the 232 edges and outside of the receiver array (Fig. 4). The filter-smoothing that we implemented decreased 233 overall error to a median of 4.4 m [0.7-27.4] (Table 2, Fig. 4B). Ten receiver stations received tag 234 signals during the mobile test and ranged from detecting 0.58% of signals to 97.87% (Fig. 5). The 235 furthest distance at which a receiver detected a signal was 14,764 m, but this was the highest of the 236 receivers (44.4 m) and only contributed to 1.17% of localisations. The receivers that detected >90% of 237 signals were within 5 km of their furthest detection. 238   for moving tags. Accuracy was higher if more receivers detected the tag. However, the tag's location 257 in respect to the receiver array configuration also had a large effect, with less accurate estimates 258 occurring when the tag was on the outskirts or outside the array of receivers. More accurate ATLAS 259 localisations can therefore be achieved through strategic placing of receiver stations. We show that 260 errors can be mitigated through a simple filter-smoothing process, as is routinely and intensively 261 applied to raw GPS location estimates (Kaplan & Hegarty, 2005). Thus, ATLAS provides a viable 262 and accurate alternative to GPS for regional-scale systems. shown to differ over short distances due to the signals being blocked by hills (Beardsworth, 2020). 275 However, the WATLAS system has very few topographical obstacles that limit reception of 276 transmissions by receivers. It is therefore likely that other factors, such as height of the tag and/or 277 receiver play an important role in reception. 278 of reflections from ground and sea, higher receiver stations can typically continue detecting tags at 281 larger distances than receivers that are closer to the earth's surface (Xia et al., 1993). This may also 282 explain why the highest receiver station (44.4 m high on Vlieland) was able to detect the tag almost 283 15 km away, but the Terschelling receiver station (35.1 m) was unable to detect the tag 10 km away. 284 While receiver height is important, tag height also affects the reception of a signal. Tags  In summary, we provide evidence for the high accuracy and precision of a relatively novel localising 339 system, ATLAS. In the design of an ATLAS system, we suggest that receiver stations are placed as 340 high as possible and surround the study site with a direct line of sight. The density of receivers is 341 important but depending on the number of receivers available for a study, trade-offs may have to be 342 made between coverage and density of receivers. Considerations should also be made according to the 343 height of the tag, with low-ground dwelling species requiring a denser receiver array than species that 344 may fly between places of interest. ATLAS provides an opportunity to track animals remotely at high 345 spatial and temporal resolution that rivals GPS technology at regional-scales and can be effective even 346 with only 3 receiver stations. While we focus our study on a flat, intertidal region, ATLAS is not 347 limited to flat landscapes. Several ATLAS systems have been operating successfully in hilly 348 landscapes and complex agricultural systems (all other ATLAS systems), as well as wetlands (Israel) 349 and woodlands (UK). These systems have successfully tracked >50 species of birds, mammals and 350 reptiles, including small (8-15g) insectivorous bats and passerine birds for which GPS tracking at high 351 temporal resolution is practically infeasible. In the focal ATLAS system, two smaller-bodied bird 352 species, sanderling, Calidris alba (~50 g) and red knot, Calidris canutus (~120g) have been 353 successfully tracked. Approximately ~200 individuals per year are tagged, enabling the monitoring of 354 complex biotic and abiotic interactions. However, it must be noted that ATLAS is limited to much 355 smaller scales than GPS and therefore is unsuitable for certain studies. Lenze Hofstee and Lydia de Loos. We would also like to thank the RV Navicula and RV Stern staff 378 and volunteers for facilitating the work around Griend and Anita Koolhaas, Hinke and Cornelis 379 Dekinga and Job ten Horn for their help with building the receiver stations. We thank Selin Ersoy, 380 who proofread the manuscript and gave constructive feedback. Finally, we thank Joah Madden and 381 Luca Borger, whose constructive conversations and comments on CEB's PhD thesis inspired the 382 writing of this manuscript.