The Argos-CLS Kalman Filter: Error Structures and State-Space Modelling Relative to Fastloc GPS Data

Abstract

Understanding how an animal utilises its surroundings requires its movements through space to be described accurately. Satellite telemetry is the only means of acquiring movement data for many species however data are prone to varying amounts of spatial error; the recent application of state-space models (SSMs) to the location estimation problem have provided a means to incorporate spatial errors when characterising animal movements. The predominant platform for collecting satellite telemetry data on free-ranging animals, Service Argos, recently provided an alternative Doppler location estimation algorithm that is purported to be more accurate and generate a greater number of locations that its predecessor. We provide a comprehensive assessment of this new estimation process performance on data from free-ranging animals relative to concurrently collected Fastloc GPS data. Additionally, we test the efficacy of three readily-available SSM in predicting the movement of two focal animals. Raw Argos location estimates generated by the new algorithm were greatly improved compared to the old system. Approximately twice as many Argos locations were derived compared to GPS on the devices used. Root Mean Square Errors (RMSE) for each optimal SSM were less than 4.25 km with some producing RMSE of less than 2.50 km. Differences in the biological plausibility of the tracks between the two focal animals used to investigate the utility of SSM highlights the importance of considering animal behaviour in movement studies. The ability to reprocess Argos data collected since 2008 with the new algorithm should permit questions of animal movement to be revisited at a finer resolution.

Fig 1. Raw Kalman filtered Argos tracks (black) and GPS tracks (red) for the focal A) bearded seal and B) ringed seal.

Dots highlight instrumentation site (blue) and the final location (yellow).

Fig 2. Spatial errors of the best modelled Argos location estimates relative to the true (GPS) position for the focal A) bearded seal and B) ringed seal using three different location error correction methods.

* denotes use of the original error structure provided the most accurate modelled locations. Red dotted lines signify the 95% percentile of the empirical cumulative distribution function (right side axis) for each suite of errors. Generally, 95% of all errors were less than 4.2 km from the true position, with modelled locations being most accurate for the ringed seal.

Fig 3. Optimal state-space modelled (SSM) Argos location data (black) overlaid with GPS locations (red) for the focal A) bearded seal and B) ringed seal.

* denotes optimal model was constructed using the error structures derived from data in [10]. Modelled location and GPS point estimates are shown as black and red dots, respectively. For each tripEstimation model, the underlying time-spent along the full path estimate is shown in purple (obscured in the bearded seal plot in favour of displaying point estimates). Black boxes highlight areas of departure by each model from the true path. Modelled location estimates for the bearded seal fitted well with GPS locations with the exception of two areas, at the very northerly edge of its trajectory, and just south of Prins Karls Forland. Note the sparse numbers of GPS location estimates for the ringed seal, particularly during transit movements between fjords. Although there were ~50% fewer Argos location estimates for the ringed seal, all models reconstructed some aspects of these transit movements despite showing a number of erroneous land locations. The exception was the tripEstimation modelled data, presumably due to the effects of incorporating a land mask during the modelling process.