Discussion
Manually sorting through movement data, visually searching for locations of mineral licks can be productive in identifying sites of interest (Poole et al 2010). However, as can be seen in Figure 3, the task can quickly become a problem of too much data, and too much complexity. To identify locations of previously unknown mineral licks, an automated method to search through potentially hundreds of thousands of data points is required.
The training data set derived from four confirmed BISH licks showed a somewhat stronger relationship with turning angle at both 125 and 250 m than when all lick sites were included (Figure 3.A-B). Although the observed difference in mean turning angle was small, with high variance, there does appear to be a relationship between high turn angles and mineral lick visits. Less clear was the relationship between step distance and lick use (Figure 3.C), though there appeared to be less variance near mineral licks locations. The predicted values from Random Forest model suggest a number of locations that could be investigated further. Interestingly, locations with higher probability values occurred as single point locations, with no nearby points showing similar increased values. The exception was the four training location sites, where multiple high valued points were present in close proximity, as would be expected. This may be the result of using the model to predict probability (similarity) for the larger movement file locations.
It should be noted that, despite the existence of many "known" mineral licks, only a handful could be confirmed from the telemetry data, and from my own limited observations. Many of the well known lick sites in the Waterton Front and Sheep River Provincial Park have been documented as used predominantly by ewes and yearling rams. It is likely that many mineral lick sites are used by specific bands of sheep. As rams are hunted in most parts of the study area for ~60 days per year, they are much more aware of human presence, and are much more easily disrupted or displaced from sites at all times of year. With increased collar deployments in March 2023 (n = 26) increased data from more individuals may help further elucidate patterns of use, however the Random Forest output of predicting each point in isolation will likely continue to limit the power of this method to accurately identify areas of interest.
To overcome this limitation, I will be continuing to explore various methods to improve predictive power. Some possibilities include incorporating a clustering algorithm to search for groups of locations (as opposed to evaluating a single point) that share similarity with points around known mineral licks, and using a weighting system to assign greater probability values to areas where multiple points share similar characteristics. I picture this as a 'moving window' type application, where instead of a linear movement of the window, it would evaluate each point in the data set by creating a buffer zone around it, and evaluate the movement data values for all points within the buffer searching for patterns. Similarly, where multiple individuals are recorded using a location at different times (or days) with a similar pattern, the area should receive greater weighting than sites where only a single individual visited a handful of times over a summer.
Despite the difficulties encountered in this study, I continue to believe these obstacles can be addressed, and a functional model can be created to simplify this, and similar, analyses.
The training data set derived from four confirmed BISH licks showed a somewhat stronger relationship with turning angle at both 125 and 250 m than when all lick sites were included (Figure 3.A-B). Although the observed difference in mean turning angle was small, with high variance, there does appear to be a relationship between high turn angles and mineral lick visits. Less clear was the relationship between step distance and lick use (Figure 3.C), though there appeared to be less variance near mineral licks locations. The predicted values from Random Forest model suggest a number of locations that could be investigated further. Interestingly, locations with higher probability values occurred as single point locations, with no nearby points showing similar increased values. The exception was the four training location sites, where multiple high valued points were present in close proximity, as would be expected. This may be the result of using the model to predict probability (similarity) for the larger movement file locations.
It should be noted that, despite the existence of many "known" mineral licks, only a handful could be confirmed from the telemetry data, and from my own limited observations. Many of the well known lick sites in the Waterton Front and Sheep River Provincial Park have been documented as used predominantly by ewes and yearling rams. It is likely that many mineral lick sites are used by specific bands of sheep. As rams are hunted in most parts of the study area for ~60 days per year, they are much more aware of human presence, and are much more easily disrupted or displaced from sites at all times of year. With increased collar deployments in March 2023 (n = 26) increased data from more individuals may help further elucidate patterns of use, however the Random Forest output of predicting each point in isolation will likely continue to limit the power of this method to accurately identify areas of interest.
To overcome this limitation, I will be continuing to explore various methods to improve predictive power. Some possibilities include incorporating a clustering algorithm to search for groups of locations (as opposed to evaluating a single point) that share similarity with points around known mineral licks, and using a weighting system to assign greater probability values to areas where multiple points share similar characteristics. I picture this as a 'moving window' type application, where instead of a linear movement of the window, it would evaluate each point in the data set by creating a buffer zone around it, and evaluate the movement data values for all points within the buffer searching for patterns. Similarly, where multiple individuals are recorded using a location at different times (or days) with a similar pattern, the area should receive greater weighting than sites where only a single individual visited a handful of times over a summer.
Despite the difficulties encountered in this study, I continue to believe these obstacles can be addressed, and a functional model can be created to simplify this, and similar, analyses.