Data and code for: A novel method for mapping high-precision animal locations using high-resolution imagery

Investigating ecological questions at the scale of individual organisms is necessary to understand and predict the biological consequences of changing environmental conditions. For small organisms this can be challenging because ecologists need tools with the appropriate accuracy, precision, and resolution to record and quantify their ecological interactions. Unfortunately, many existing tools are only appropriate for medium to large organisms or those that are wide ranging, inhibiting our ability to investigate the spatial ecology of small organisms at fine scales. Here, we tested a novel workflow for recording animal locations at very fine (decimeter) spatial scales, which we refer to as High-resolution Orthomosaic Location Recording (HOLR). The workflow for HOLR combined direct observations with data collection of locations on high-resolution uncrewed aerial vehicle (UAV) imagery loaded on smartphones. Observers identified landscape features they recognized in the imagery and estimated positions relative to these visual landmarks. We found HOLR was approximately twice as accurate as consumer-grade GPS devices, with a mean error of 0.75 m and a median error 0.30 m. We also found that performance varied across landscape features, with the highest accuracy in areas that had more visual landmarks for observers to use as reference points. In addition to sub-meter accuracy, HOLR was cost-effective and practical in the field, requiring no bulky equipment and allowing observers to easily record locations away from their own location. This workflow can be used to record locations in a variety of situations, but it will be particularly cost-effective when users simultaneously utilize the high-resolution environmental data contained within UAV imagery. Together, these tools can expand the application of spatial ecology research to smaller organisms than ever before.

Methods
Target Deployment
We deployed 2.5 cm x 15 cm painted copper pipes across the study site. These targets were operative thermal models (OTMs) that were primarily being deployed to measure temperature across the study site as part of a separate study, but they made suitable targets because the pipes were designed to approximate the size and appearance of a small, elongate animal, such as a lizard. We deployed approximately 28 targets in each of four sessions, each lasting roughly 4 weeks (total n = 111). We used a stratified random design to ensure the targets were located in or around a variety of landscape features present at the site, including on D. ingens burrow mounds (n = 63), off D. ingens burrow mounds (n = 12), under shrubs (n = 24), and in a wash (n = 12; Figure 5). The off-burrow feature types had few easily identifiable visual landmarks that observers could use while mapping locations (Figure 6a and b). The on-burrow feature type had some potential landmarks, including burrow entrances and disturbed dirt, but these were sometimes very similar to one another, reducing their utility as visual landmarks (Figure 6a, b and e). The shrub feature type always had at least one shrub present, but often more than one shrub was nearby, providing a number of landmarks to use when mapping (Figure 6c and e). Finally, the wash feature type had many visual landmarks, including shrubs, large rocks, and clear banks and ledges (Figure 6d and e). While most of the targets used were placed individually, the wash sites were placed in four groups of three (one group per session), with one target in the middle of the wash channel and one on each bank to better capture the effect of these topographic changes on temperature for a separate study. These targets were far enough apart that recording the location of one did not influence where an observer would record the other wash targets, so they were treated as independent targets. We used the Eos Arrow Gold RTK GNSS receiver to record the locations of all targets. All locations were recorded in ArcGIS Collector, averaging across 30 locations, and had an estimated error of 3 cm or less.

Target Mapping
Two observers recorded the locations of all targets at the beginning of each of the four sessions. This design allowed for the comparison of accuracy between observers and through time. Observers worked independently, without communicating about visual references around target locations during the survey. One observer had extensive prior experience using the Avenza app and a small amount of experience mapping onto UAV imagery from a pilot study. The other observer only had a brief training session on the app prior to this work, but otherwise had the education and experience typical of an ecology research assistant. We used this difference in experience to investigate the influence of experience on accuracy. In addition to mapping the target locations every four weeks (i.e., mapping sessions 1-4), both observers used the same HOLR method to regularly record locations of radio-collared G. sila at the study site, and so were gaining mapping experience for the duration of the study. The observers treated the targets as if they were live animals, staying at least 8 m away from the targets while recording their locations. It is important to avoid disturbing study organisms when collecting animal location data to avoid influencing their behavior or the data collected.

            To provide a comparison to the most common method for recording locations in the field, we also recorded the location of each target using a Garmin eTrex Vista handheld GPS receiver for the first three sessions. During the third session this unit failed, so we switched to an eTrex Legend H (Garmin Ltd., Olathe, KS, USA) for the remainder of session 3 and all of session 4, such that we recorded 58 locations with the Vista and 53 with the Legend H. All locations were collected with the GPS receiver approximately two meters above the ground, clear of any obstructions. The receiver was held horizontally in this position for approximately ten seconds before collecting the location as a single point. We chose this approach to balance the desire for optimal GPS performance with the time constraints a researcher typically faces when tracking animals in the field. Full details on the HOLR workflow are available in supplemental information (Appendix S1).

Accuracy Analysis
Due to low sample sizes, we used summary statistics and visual assessments of the data to compare the accuracy of the GPS receivers against the two observers using HOLR. We used linear models and AIC scores to assess the influence of three variables on HOLR accuracy (Burnham and Anderson 2002). The distribution of error values was right skewed, so we used the natural log of error as the response variable for the models (Appendix S2). To explore how the presence of different landscape features might affect HOLR users’ performance, we included a categorical variable of landscape feature type with four levels (i.e., shrubs, wash, on-burrows, and off-burrows). We included observer as a categorical variable with two levels, to control for any differences in performance. Finally, to assess changes in performance over the duration of the study, we included session as a continuous variable. We created seven total models, using each unique combination of the three predictor variables. We used AIC scores to rank the models against each other, with lower scores indicating better models (Burnham and Anderson 2002).

Other Processing Notes:

The spatial data (Avenza, GNSS and GPS) is not processed other than the GNSS and GPS files merged across several collecting sessions. The OTM GPS Locations.csv file combines the location, target, session and device used for all GPS locations. The OTM Error Independent Long.csv file has the error for all mapped and GPS locations calculated in R, with targets that were close to another (i.e., not independent) removed and the data arranged in long format. See included R scripts for more processing details.