- Top of page
- Study area
The first type of error inherent to GPS technology, unsuccessful fix acquisition, leads to missing location data (D’Eon et al. 2002; Frair et al. 2004). Fix rates for stationary GPS collars can range from 68% to 100%, with most collars achieving rates > 85%; however, fix rates can be as low as 13% (Frair et al. 2004). Environmental factors that can affect reception of GPS satellite signals and result in failed location acquisition include a high percentage of tree canopy cover (Rempel & Rodgers 1997; Obbard, Pond & Perera 1998; D’Eon et al. 2002; Di Orio, Callas & Schaefer 2003; Frair et al. 2004), cover types (Di Orio, Callas & Schaefer 2003), the interaction between canopy cover and terrain (D’Eon et al. 2002; Frair et al. 2004), time of day (Moen, Pastor & Cohen 1997; Frair et al. 2004; but see D’Eon et al. 2002) and time of year (Dussault et al. 2001; Frair et al. 2004). Also, in the northern hemisphere, GPS collars can have higher fix rates on south-facing slopes compared with north-facing slopes (D’Eon & Delparte 2005). Therefore GPS collar location data may be biased towards acquiring satellite fixes in more open habitats with favourable topography. In addition, animal behaviour may dramatically affect GPS collar fix rates (Dussault et al. 2001; D’Eon 2003; D’Eon & Delparte 2005). Collar location acquisition schedules can also influence fix rates, with higher fix rates observed for collars that attempt location fixes at shorter time intervals (Cain et al. 2005). To counter biases associated with missed locations, correction factors may be developed and applied to GPS location data sets (Frair et al. 2004).
The second error type, location error, is inherent in all telemetry systems and can lead to misclassification of habitats used in studies of resource selection (Rodgers 2001) and bias estimates of movement paths (DeCesare, Squires & Kolbe 2005). Location error is largely influenced by habitat components, such as canopy cover (D’Eon et al. 2002; Di Orio, Callas & Schaefer 2003) and terrain obstruction (D’Eon et al. 2002; Cain et al. 2005), although atmospheric conditions can also contribute. As selective availability (a USA government strategy imposed on the collection of location data to make locations unpredictably inaccurate for national defence concerns) was discontinued in May 2000, most stationary test locations are reported to be accurate to within 30 m (D’Eon et al. 2002) and can be accurate to < 3 m (Rodgers 2001). Two-dimensional (2-D) and three-dimensional (3-D) fixes are recorded when the GPS unit simultaneously contacts three and ≥ 4 satellites, respectively, with 3-D fixes generally more accurate than 2-D fixes (Moen et al. 1996; Rempel & Rodgers 1997; Di Orio, Callas & Schaefer 2003). Increasing canopy cover may result in fewer 3-D fixes and more 2-D fixes as a result of satellite signal obstruction (Rempel, Rodgers & Abraham 1995).
A PDOP value, which measures satellite geometry, is recorded for each GPS location. Lower PDOP readings indicate wider satellite spacing, which potentially minimizes triangulation error and results in more accurate location estimates (D’Eon et al. 2002; D’Eon & Delparte 2005). Because PDOP is related to location error (D’Eon et al. 2002), screening out location data with high PDOP values has been suggested to reduce location error (Moen et al. 1996; Dussault et al. 2001; D’Eon & Delparte 2005). Also, the screening of 2-D fixes from data sets may substantially decrease location error (Moen et al. 1996; D’Eon et al. 2002). While screening data sets can reduce location errors, it also may lead to significant reductions of location data and introduce additional biases into analyses of animal locations. More detailed investigation of data screening is necessary to evaluate the trade-offs between eliminating inaccurate locations and retaining the maximum amount of location data. Our objectives were to quantify how habitat affected GPS collar performance across the range of variation within a mountainous region of northern Idaho, USA, and evaluate different data screening options. Specifically, we assessed: (i) how canopy cover and terrain obstruction simultaneously affected fix rate, proportion of 3-D fixes, location error and PDOP values; (ii) the relationship between PDOP values and location error; and (iii) how four data screening options affected location error and data reduction for GPS location data sets from stationary test collars and free-ranging black bears Ursus americanus.
- Top of page
- Study area
The study area was located in the Purcell Mountains of northern Idaho, USA, and was bordered to the west by the Kootenai River, north by Canada, east by Montana, and south by Highway 2. The area was approximately 300 km2 in size. Highway 95 ran through the centre of the study area and an extensive network of secondary roads was present because of timber harvest. The topography was varied, ranging from flat valley bottoms to steep and rugged mountain slopes, with altitudes ranging from 700 m to > 1900 m. The regional climate was Pacific maritime, with cold, snowy winters and short, warm summers, with annual precipitation averaging 100–150 cm, mostly in the form of snow (Kasworm, Carriles & Radandt 2004). Forest predominated throughout the study area, although extensive open areas were also present in meadow sites and recently logged timber stands. At drier, lower altitudes, common tree species included ponderosa pine Pinus ponderosa, western larch Larix occidentalis and Douglas fir Pseudotsuga menziesii. More mesic, low-altitude stands consisted of grand fir Abies grandis, western red cedar Thuja plicata, western hemlock Tsuga heterophylla and black cottonwood Populus trichocarpa. Subalpine fir Abies lasiocarpa, Engelmann spruce Picea engelmannii and mountain hemlock Tsuga mertensiana occurred at higher altitudes above 1500 m, and lodgepole pine Pinus contorta and western larch were common at mid-altitudes.
- Top of page
- Study area
The overall fix success for Lotek 3300 L GPS collars placed at test locations was high for all 16 habitat categories (mean 99·5%, SE 0·14, range 97·9–100%). Fix rate was not explained well by models that included habitat variables, canopy cover and satellite view (Table 1). However, missed fixes tended to occur in habitats with low satellite view, high canopy cover or a combination of both.
Table 1. Models, with associated R2, AIC, ΔAIC values, and model parameter estimates evaluating how habitat variables affected overall fix rate, 3-D fix rate, location error and PDOP values for GPS collars at 48 test sites in the Purcell Mountains, Idaho, USA, 2005.
|Dependent variable*||Model variables†||R2||AIC||ΔAIC||Model parameter estimates|
|Fix rate (FR)||SV||0·065||114·121||0||FR = 77·53 + 12·02 SV|
|SV CC||0·096||114·536||0·415||FR = 78·34 + 12·2 SV − 1·8 8 CC|
|CC||0·028||115·976||1·855||FR =−1·82 + 89·36 CC|
|SV CC SV × CC||0·099||116·334||2·213||FR = 73·90 + 13·17 SV + 6·68 CC − 9·40 SV × CC|
|3-D fix (3D)||SV CC||0·732||184·546||0||3D = 31·42 + 58·31 SV − 32·84 CC|
|SV CC SV × CC||0·742||184·687||0·141||3D = 3·74 + 88·72 SV + 20·60 CC − 58·66 SV × CC|
|CC||0·626||198·511||13·965||3D = 84·10 − 32·54 CC|
|SV||0·095||240·936||56·390||3D = 17·32 + 55·07 SV|
|Location error (LE)||SV CC||0·582||254·820||0||LE = 61·26 − 62·46 SV + 50·00 CC|
|SV CC SV × CC||0·582||255·819||0·999||LE = 62·76 − 64·11 SV + 47·10 CC + 3·18 SV × CC|
|CC||0·537||256·696||1·876||LE = 4·82 + 49·68 CC|
|SV||0·038||291·801||36·981||LE = 82·73 − 57·54 SV|
|PDOP (PD)||SV CC SV × CC||0·519||−12·784||0||PD = 16·88 − 14·54 SV − 16·06 CC + 20·05 SV × CC|
|SV CC||0·383||−2·855||9·929||PD = 7·42 − 4·15 SV + 2·21 CC|
|CC||0·322||−0·325||12·459||PD = 3·67 + 2·18 CC|
|SV||0·055||15·599||28·383||PD = 8·37 – 3·93 SV|
Both canopy cover and satellite view were good predictors of the proportion of 3-D fixes, mean location error and mean PDOP values at the 48 test sites (Table 1). In all the top models, canopy cover contributed the greatest to model prediction (based on partial R2 values), although satellite view and the interaction of canopy cover and satellite view improved the models. As canopy cover increased and satellite view decreased, the proportion of 3-D fixes decreased, location error increased and PDOP values tended to increase.
The mean location error for all individual test locations sampled across the range of habitat variation was 14·3 m (SD 3·18 m), with a 95% CEP of 106·8 m (Table 2). Mean location error and CEPs for the total data set varied substantially within different habitat groups. For example, the 95% CEP increased from 19·9 m to 284·3 m when comparing sites with low canopy cover and low terrain obstruction to high canopy cover and high terrain obstruction; in addition, within these two groups the percentage of 3-D fixes decreased from 97·7% to 58·9%, respectively (Table 2). Furthermore, within signal quality categories (3-D, 2-D, PDOP < 10, PDOP < 6 and PDOP < 3), larger errors were observed as satellite view decreased and canopy cover increased. For example, for locations with PDOP < 10, the 95% CEP decreased from 246·0 m to 15·5 m when comparing sites with high canopy cover and high terrain obstruction to low canopy cover and low terrain obstruction (Table 2). Location errors for 2-D fixes (mean 36·0, SD 3·24, n= 613) were consistently larger than 3-D fixes (mean 11·8, SD 2·86, n= 2828) for all habitat categories (Table 2; F1,44 = 7·97, P= 0·0071).
Table 2. GPS collar location errors based on location type and habitat for GPS collars placed at 48 test sites in the Purcell Mountains, Idaho, USA, 2005. For each location type, the first line includes all locations, with the second line representing the best satellite reception conditions, the third and fourth lines representing intermediate satellite reception conditions, and the last line representing the poorest satellite reception conditions.
|Location type||Habitat* (SV–CC)||Locations in group (%)||Location error(m)|
| ||All||100||14·3 (3·18)||13·8||106·8||262·6||942·8|
|1–1||100|| 5·0 (2·56)|| 5·0|| 19·9|| 90·2||137·9|
|4–1||100||13·6 (2·58)||12·4|| 74·5||127·1||367·5|
| ||All|| 82·2||11·8 (2·86)||11·7|| 68·1||131·6||367·5|
|1–1|| 97·7|| 4·8 (2·50)|| 4·9|| 18·0|| 65·2||137·9|
|1–4|| 74·9||18·7 (2·37)||17·8|| 82·3||187·8||318·0|
|4–1|| 87·4||12·9 (2·52)||11·2|| 71·1||131·6||367·5|
|4–4|| 58·9||36·7 (2·41)||46·8||126·3||344·6||344·6|
| ||All|| 17·8||36·0 (3·24)||35·1||253·0||565·9||942·8|
|1–1|| 2·3||18·0 (3·28)||14·3|| 90·2|| 90·2|| 90·2|
|1–4|| 25·1||50·7 (3·46)||48·6||364·5||772·1||772·1|
|4–1|| 12·6||19·1 (2·88)||15·1|| 74·5|| 76·1|| 76·1|
|4–4|| 41·1||65·7 (2·84)||57·4||557·0||654·6||654·6|
|PDOP < 10|
| ||All|| 92·5||12·8 (2·97)||12·6|| 78·3||201·7||942·8|
|1–1|| 97·2|| 4·7 (2·40)|| 4·9|| 15·5|| 38·7|| 90·2|
|1–4|| 89·3||20·1 (2·54)||18·6||114·3||261·3||364·5|
|4–1|| 90·7||11·9 (2·40)||10·8|| 60·0||105·3||367·5|
|4–4|| 85·1||41·2 (2·58)||50·5||246·0||557·0||587·9|
|PDOP < 6|
| ||All|| 81·4||11·7 (2·88)||11·7|| 68·6||147·4||587·9|
|1–1|| 93·5|| 4·6 (2·39)|| 4·8|| 15·0|| 38·7|| 90·2|
|1–4|| 79·1||18·1 (2·32)||17·7|| 77·8||143·0||261·3|
|4–1|| 78·6||11·4 (2·31)||10·6|| 58·6|| 83·4||105·3|
|4–4|| 70·2||38·9 (2·67)||48·9||314·9||587·9||587·9|
|PDOP < 3|
| ||All|| 32·1|| 9·5 (3·04)|| 8·9|| 65·5||138·9||383·3|
|1–1|| 59·7|| 4·2 (2·24)|| 4·5|| 14·3|| 29·1|| 32·7|
|1–4|| 21·9||19·9 (2·44)||18·8||101·0||143·0||143·0|
|4–1|| 24·7||10·5 (2·19)|| 9·6|| 48·6|| 74·5|| 74·5|
|4–4|| 30·5||35·8 (2·46)||45·5|| 95·4||321·9||321·9|
The relationship between PDOP and location error varied by habitat category and signal quality. There was a strong positive relationship between mean PDOP and mean location error at test sites (Fig. 1; R2 = 0·779, n= 48) and a weak relationship between PDOP and location error for all individual locations (R2 = 0·143, n= 3441). When individual location errors and PDOP values were evaluated based on signal quality (2-D vs. 3-D), the R2 values increased slightly and revealed differing slopes (Fig. 2). Also, a greater number of large outlier location errors were observed for 2-D locations compared with 3-D locations.
Figure 1. Relationship between mean LN location error and mean PDOP for 48 test sites in the Purcell Mountains, Idaho, USA, 2005.
Download figure to PowerPoint
Figure 2. Relationship between location error and PDOP values for all individual (a) 2-D locations (n = 613) and (b) 3-D locations (n = 2828) in the Purcell Mountains, Idaho, USA, 2005.
Download figure to PowerPoint
The application of the four data screening options to GPS location data sets from test collars and free-ranging black bears resulted in varying amounts of data reduction and accuracy. For the test collar data set, the first option (screen all locations with PDOP > 10) resulted in the elimination of 7·5% of locations, failed to purge several large location errors, and resulted in a 95% CEP of 78·3 m (Table 3). Under the second option (screen 2-D fixes with PDOP > 5), the majority of large outlier location errors were purged from the data set, 4·5% of locations were eliminated, and the 95% CEP equalled 79·2 m. The third option (screen 3-D fixes with PDOP > 10 and 2-D fixes with PDOP > 5) also removed most large outlier locations, eliminated 10·4% of locations and resulted in a 95% CEP of 69·4 m. The fourth option (screen all 2-D fixes) removed the most large location errors and eliminated 17·8% of locations, with the 95% CEP equal to 68·1 m. The four data screening options resulted in similar mean location errors for the test collar data set, ranging from 11·8 m to 13·2 m (Table 3). However, the percentage of accurate locations (location errors ≤ 100 m) eliminated varied substantially from 2·0% under option 2 to 14·9% under option 4 (Table 3). The percentage of all locations with errors > 100 m that were eliminated from the data set for options 1, 2, 3 and 4 were 38·7%, 43·3%, 61·3% and 67·0%, respectively. Elimination of large errors (> 300 m) followed a similar pattern (Table 3).
Table 3. Data reduction and location accuracy resulting from four data-screening options for 48 test collar sites in the Purcell Mountains, Idaho, USA, 2005.
|Data-Screening options||Total data retention (%)||% of eliminated locations with errors > 100 m||% of all locations with errors > 300 m eliminated||% of all locations with errors 100 m eliminated||Location error (m)|
|Mean error (SD)*||50%†||95%†||99%†||100%†|
|Original data||99·5||NA||NA||NA||14·3 (3·18)||13·8||106·8||262·6||942·8|
|1 (all PDOP > 10)||92·5||30·1||51·9|| 5·3||12·8 (2·97)||12·6|| 78·3||201·7||942·8|
|2 (2-D PDOP > 5)||95·5||56·4||63·0|| 2·0||13·2 (2·96)||13·0|| 79·2||164·0||587·8|
|3 (3-D PDOP > 10, 2-D PDOP > 5)||89·6||34·5||70·4|| 7·0||12·2 (2·85)||12·2|| 69·4||138·5||587·8|
|4 (all 2-D)||82·2||21·2||85·2||14·9||11·8 (2·86)||11·7|| 68·1||131·6||367·5|
The GPS collar fix rate for 10 free-ranging black bears averaged 92·1%, ranging from 88·7% to 95·8% (Table 4). The mean fix rate for black bear GPS collars was 7·3% lower than for stationary test collars placed in forest with > 50% canopy cover, indicating that animal behaviour affected fix rate. For black bear data sets, data screening options 1 and 2 resulted in similar amounts of total data retention, 84·6% and 84·2%, respectively, with options 3 and 4 exhibiting greater data reduction, of 80·0% and 61·0%, respectively (Table 4).
Table 4. Reduction of location data resulting from four data-screening options for 10 black bear GPS collar data sets in the Purcell Mountains, Idaho, USA, 2005.
|Black bear individuals||GPS collar fix rate (%)||Data screening options|
|Option 1 All PDOP > 10||Option 2 2-D PDOP > 5||Option 3 3-D PDOP > 10, 2-D PDOP > 5||Option 4 All 2-D|
|Screening* data loss (%)||Total† data retention (%)||Screening* data loss (%)||Total† data retention (%)||Screening* data loss (%)||Total† data retention (%)||Screening* data loss (%)||Total† data retention (%)|
|Mean (SD)||91·8 (2·39)||8·1 (0·63)||84·3 (2·38)|| 8·6 (1·22)||83·9 (3·09)||13·3 (1·21)||79·6 (2·83)||34·8 (5·06)||59·9 (5·85)|
- Top of page
- Study area
The GPS collar fix rate for test collars was close to 100% within all habitats in our study area. Although the fix rate was high, missed fixes most often occurred at test sites with relatively high terrain obstruction, high canopy cover or a combination of both. Therefore, study areas with greater terrain obstruction and consistently high canopy cover may result in a higher proportion of missed fixes. In addition, greater terrain obstruction could potentially lead to greater location error because collars would receive location signals from, on average, fewer satellites that exhibit poorer satellite configuration, thus increasing PDOP values and the number of 2-D fixes.
This is the first study to report how location errors and PDOP values are affected by continuous ranges of canopy cover and terrain obstruction simultaneously. Our work expands on that of D’Eon & Delparte (2005), who evaluated the relationship between location error and PDOP values for GPS collars in the Selkirk Mountains of south-eastern British Columbia. The authors placed GPS collars in an open, grassy field with 0% canopy cover and little terrain obstruction and recorded a 95% CEP of 28·9 m; however, they acknowledged that higher canopy cover and terrain obstruction values may further influence location error. Our results demonstrate that both of these habitat variables can contribute significantly to PDOP values and location error. We report a 95% CEP of 106·8 m for all our locations, which sampled a range of terrain obstruction and canopy cover values. When evaluating our habitat category with the lowest terrain obstruction and lowest canopy cover, the 95% CEP equalled 19·9 m, which is comparable to D’Eon & Delparte's (2005) study.
Our GPS collar tests revealed more outlier location errors than previously reported. From 3441 test locations, we recorded 27 locations with errors > 300 m, compared to two outliers with errors > 300 m out of 6359 locations (D’Eon & Delparte 2005). These results can probably be attributed to our study sampling a range of canopy cover and terrain obstruction values. Although GPS collar performance can vary considerably among collar manufacturers (Frair et al. 2004), we believe that our results using Lotek collars are comparable with D’Eon & Delparte's (2005) study, which used Advanced Telemetry System (ATS, Isanti, MN) GPS collars, because previous research has shown Lotek and ATS collars to exhibit similar fix rates and location errors (Di Orio, Callas & Schaefer 2003; Frair et al. 2004).
Our study demonstrates that location errors, and possibly fix rates, for GPS collars should be quantified and reported in relation to specific habitats. For example, when evaluating location error the 95% CEP for the entire test collar data set ranged from 19·9 m to 284·3 m between the best and worst habitat categories for satellite reception, respectively. Evaluation of location error in areas with very low canopy cover and little terrain obstruction should be considered ideal conditions; extrapolation of results from tests conducted at such sites to habitats with higher canopy cover and terrain obstruction is inappropriate and will underestimate true location errors.
To screen GPS location data, a PDOP cut-off may effectively remove large location errors and subsequently increase accuracy of location data sets (Moen et al. 1996; Dussault et al. 2001; D’Eon & Delparte 2005). Two-dimensional fixes are less accurate than 3-D fixes, and demonstrate differing slopes when related to PDOP values (Fig. 2; Moen et al. 1996); therefore, it is appropriate to treat these two categories individually for screening. When screening data, the mean location error generally is not greatly improved; however, large location errors may effectively be eliminated from data sets, often decreasing CEP values (Tables 2 and 3; D’Eon & Delparte 2005).
When screening location data, there is a trade-off between data reduction and data accuracy. For our first option (screening all data collectively), a PDOP cut-off of 10 has been recommended to improve location accuracy and remove unusually large location errors, while only eliminating 1·3% of original locations (D’Eon & Delparte 2005). However, using this recommendation with our test collar and black bear data sets we eliminated 7·5% and 8·1% of locations, respectively, and still retained several large location errors. While the mean location error was similar for all screening options, the CEPs were most affected by screening 2-D fixes (options 2, 3 and 4). Because the majority of large location errors were associated with 2-D fixes with PDOP values > 5 (Fig. 2), removing these fix types resulted in the elimination of most outlier location errors and reduced CEP. However, data reduction varied considerably for options screening 2-D fixes, ranging from 8·6% to 34·8% for black bear data sets.
We believe that the most effective data screening approach is to focus on screening out 2-D locations at a specific PDOP cut-off. For our data sets, the second option we evaluated (screening 2-D locations with a PDOP > 5) was a suitable compromise between reducing large location errors and minimizing data reduction. The second option was most efficient in eliminating location errors > 100 m, eliminated only a small fraction of locations with errors ≤ 100 m, and resulted in total data retention of about 84% for black bear data sets. To increase the accuracy of location data further, a reasonable consideration would entail eliminating 3-D locations with high PDOP values (option 3), although data loss may be relatively large. Removing all 2-D locations resulted in the most accurate data set; however, data reduction was also the most substantial and probably resulted in an unjustifiable amount of data reduction. To determine an appropriate PDOP threshold value for data screening, the proportion of locations with relatively large errors could be evaluated across the range of PDOP values to assist in identifying acceptable levels of location error and data reduction.
Misplaced locations can bias analyses of habitat selection and space use, which has been evaluated in VHF telemetry (Nams 1989; Samuel & Kenow 1992; Withey, Bloxton & Marzluff 2001) and should also be addressed for GPS telemetry locations. To reduce the bias in identifying areas used by animals for habitat analyses, buffers could be applied to locations that are related to the magnitude of the location error (Rettie & McLoughlin 1999). Because PDOP values are related to location error for GPS telemetry, they could be used to set buffer sizes. In addition, while data screening may effectively remove large location errors from GPS telemetry data sets, it could potentially introduce biases that affect estimates of habitat and space use by eliminating locations associated with habitats that induce greater errors. Correction factors may be developed and applied to screened data sets to address this potential bias. Studies have examined how to correct for biases associated with missed GPS fixes based on habitat characteristics and space use (Frair et al. 2004; Horne, Garton & Sager 2007) and similar approaches may be appropriate for data sets that are screened to remove large location errors.
As canopy cover and terrain obstruction increases location data are more likely to exhibit higher PDOP values and fewer 3-D fixes, both of which indicate greater location error. Therefore the performance of GPS collars is largely dependent on the habitats that collared animals use. For example, greater location errors would be expected for animals that spend time in closed-canopy forests compared with more open environments. In addition to habitat variables, animal behaviour can significantly affect GPS collar performance. Our study has demonstrated that animal behaviour can reduce fix rates by as much as 11%. Future research evaluating relationships between GPS collar performance, animal behaviour and whether specific animal activities occur in different habitats could be used to develop correction factors for missed fixes and location errors in relation to animal behaviour and habitat selection.
- Top of page
- Study area
Funding and support for this project were provided by the Idaho Department of Transportation, Idaho Department of Fish and Game, US Forest Service and the University of Idaho. We thank E. Strand for assistance with GIS analysis. We also thank R. D’Eon, J. Frair, H. Beyer, G. Smith, K. Sager and R. Hoffman for discussions and assistance regarding the evaluation of GPS radio collar performance. W. Wakkinen, J. Hayden, T. Johnson, T. Radandt and W. Kasworm assisted with field work. W. Estes-Zumpf, R. Long, J. Muir, D. Sanchez, W. Wakkinen, P. Zager and two anonymous referees provided helpful comments on this manuscript.