Improvements on GPS Location Cluster Analysis for the Prediction of Large Carnivore Feeding Activities: Ground-Truth Detection Probability and Inclusion of Activity Sensor Measures

1. What they did

Researchers often find where and when large carnivores feed by looking for clusters of Global Positioning System (GPS) collar locations, then sending people to some of those clusters to look for prey remains (“ground-truthing”). The results are used to build a model that predicts which of the unvisited clusters are feeding sites. The authors, working on cougars in the Front Range of Colorado, tested two ways to improve these models.

• Collars. 33 cougars wore GPS collars that took 7 locations a day (every 3 hours at night, every 4 hours by day). The collars also had a two-axis activity sensor (accelerometer): one axis measured forward and backward movements, the other side-to-side and head-rolling movements, summed every 288 seconds.
• Choosing clusters. Each month for three years (2010 to 2012), a computer script grouped the locations into clusters. Two clusters of at least 2 locations within 200 m and 4 days, and one cluster of two locations 200 to 500 m apart with a missed location between them, were picked at random each month for a visit.
• Searching. Observers visited each cluster in the following month, on average 29.6 days after the cougar arrived. They searched a 50 m radius around each GPS location until they found prey remains, and recorded prey species, age and sex. Any carcass counted as feeding, whether or not the cougar had killed it. New observers were trained alongside an experienced observer for 2 to 4 weeks, and observers rotated between cougars and areas each month.
• Modelling. They predicted presence or absence of prey remains from four groups of measures: the pattern of GPS locations (how many, what share at night, how spread out); the activity sensor readings; ground-truthing error measures (search delay, share of the cluster’s locations the observer had in hand, GPS fix success); and season. They compared 16 models and applied them to all 12,096 clusters, with and without correcting for ground-truthing errors.
• Double-observer test. Some clusters were visited soon after the cougar started feeding (from about 6 hours up to 2 days after the cluster ended), using locations sent by satellite in near real time. When one of these clusters was later picked at random for a standard visit, the second observer did not know it had already been visited. The authors compared the two observers’ results.

2. What they found

Data collected

• The 33 cougars were followed for an average of 495 days each, giving 90,950 usable GPS locations, 12,096 clusters and about 5.6 million activity readings per axis.
• Observers visited 1,171 clusters. Prey ranged from 1 to 450 kg, and 38.2% were small prey (under 20 kg).

The prediction model

• The best model used all four groups of measures. It predicted feeding with an AUC (a 0 to 1 score of how well the model tells feeding from non-feeding clusters) of 0.893, or about 90% accuracy.
• Feeding was more likely when more locations were at night, when there were more locations, and when the cluster was tightly grouped. Many locations with few at night tended to be day beds. A spread-out cluster could still be a feeding site when few locations were at night, for example when a cougar bedded away from the carcass but within 200 m.
• Activity sensors helped most. Activity alone predicted almost as well as the GPS pattern alone (AUC 0.819 versus 0.823). Together they raised performance by about six percentage points. Feeding was more likely when forward and backward movement was strong compared with side-to-side movement; the authors presume feeding involves more forward and backward motion and travelling more side-to-side motion, but had no direct observations to confirm this.
• Ground-truthing errors. Feeding was found less often the longer the search delay (a simple straight-line decline) and when observers had fewer of the cluster’s GPS locations in hand. Adding these measures improved model fit but hardly changed prediction. Correcting for them raised the predicted number of feeding events from 3045 to 3360, about 10% more.
• Season. Feeding was more likely in summer.
• Adding each cougar as a random effect did not improve the models.

Double-observer test

• 174 clusters were visited twice. Of these, 65.5% held large prey (adult ungulates), 16.1% small ungulates (fawns), 12.1% non-ungulates (such as mesocarnivores), and 6.3% had no prey remains.
• The two observers disagreed on 6 clusters (3.5%). The standard (later) observer missed prey remains twice, once small prey. Surprisingly, the early observer missed remains 4 times, 3 of them small or non-ungulate prey. In all, 4 of the 6 missed sites were small prey.
• The early observers often had only the first two GPS locations of the cluster and had to avoid disturbing the cougar, so their searches were less thorough. In about 25% of early visits the cougar was near the cluster; most stayed close to the carcass, and those that moved off came back soon after or later that evening.

3. What they concluded

• Adding activity sensor data to the standard GPS measures clearly improves predictions of where and when large carnivores feed, and the simple, coarse sensors used here are available to most researchers.
• Ground-truthing misses only a small share of feeding events (about 4%), even with an average search delay of about 30 days. Missed events depend on the delay and on how many of the cluster’s GPS locations the observer has in hand, and small prey are missed most.
• This is the first study to correct kill counts for ground-truthing errors. Because many earlier cougar studies used similar methods, their kill rates may have been biased low.
• Future studies should give observers all of a cluster’s GPS locations before they search, use shorter GPS intervals where battery life allows, and develop methods that combine GPS and activity data in one step. Visiting kills early is often the only way to tell whether the predator killed or scavenged, but it risks missing remains and disturbing the predator.

4. How tracking was or was not used

Tracks and signs identification was central to ground-truthing. Observers searched a 50 m radius around each GPS location for prey remains, decided whether a feeding event had taken place, and identified prey species, age and sex. These field results were the “truth” that every prediction model was built on. The double-observer test measured how often feeding signs were missed. No trailing is reported, and the paper does not say which signs observers used to identify prey or to judge the cause of death.

5. Tracker or observer certification

Reported but not validated. Novice observers were trained and accompanied by an experienced observer for 2 to 4 weeks before searching alone, to develop a “search image” for prey remains, and observers were rotated monthly to spread differences in ability. No certification or individual skill test is reported. The double-observer test measured agreement between two visits (96.5%), not each observer’s skill against a known answer. Fourteen people who carried out ground-truthing visits are thanked in the acknowledgments.

6. Other technologies, skills or innovations

• Vectronic GPS-Plus collars with Globalstar or Iridium satellite upload of locations in near real time, on-board memory, and a two-axis accelerometer activity sensor (readings from 0 to 255 per axis every 288 seconds), downloaded from the ground by VHF (very high frequency) radio or when the collar was removed.
• Activity readings averaged over 1.5 hours around each GPS location, including the difference between the two axes, which helped tell feeding from other activity.
• A randomized, month-by-month sampling design for choosing clusters to visit, which the authors say was critical for measuring the effect of search delay.
• A blind double-observer design to measure missed feeding events.

7. Implications for conservation (from the paper)

• Kill rates estimated from GPS clusters in earlier studies may be too low, which matters for predation estimates.
• Early visits to kills risk disturbing the predator; the authors suspected a few kills were abandoned but could not say whether this was caused by observers.