Advancing spatial analysis of invasive species movement data to improve monitoring, control programs and decision making: feral cat home range as a case study

Introduction

Invasive animals are a significant global threat to biodiversity (Doherty et al. 2016), primary production (Schley et al. 2008) and human health (Chinchio et al. 2020). Many invasive animals are active over large areas (e.g. feral pig), or else are solitary and elusive (e.g. feral cat). These challenges make monitoring and control programs resource hungry and if not planned well, ineffective. Density and abundance are often determined by the placement of monitoring devices at defined intervals (e.g. Allen et al. 2013) or else at random (e.g. Guerrasio et al. 2022). However, in these situations, target species detection is reliant upon random encounters with monitoring devices. Similarly, large numbers of poison baits are often scattered broadly with the intent to saturate the landscape to increase the likelihood of bait encounters (e.g. Fancourt et al. 2021). In Fancourt et al. (2021)’s case, bait placement along tracks proved insufficient to significantly reduce the feral cat population, highlighting that in such cases, an over-reliance upon random encounters is likely to be both ineffective and inefficient. This was not the case in Ballard et al. (2020), where landscape saturation of baits resulted in a 90% mortality rate in collared wild dogs. However, such saturation may also increase the likelihood of nontarget species mortality. As such, improving monitoring or control tool placement to target locations of likely higher use of the intended species (Harriott et al. 2021; Wilson et al. 2023a), is likely to improve encounter rates in monitoring and control programs, while simultaneously reducing the probability of nontarget encounters.

Home range areas are widely reported in invasive animal literature (Amos et al. 2014; McGregor et al. 2015; Wilson et al. 2023b) but have limited practical value for improving monitoring or control programs. Animal-borne global positioning system (GPS) units are often used to collect spatial information about an animal’s movements, with location fixes being recorded by the GPS at predetermined time intervals. An animal’s home range calculation typically encompasses 90–100% of the location fixes, but in doing so, it fails to identify key areas of higher use. Most animals do not use the landscape uniformly, but instead have core areas where they spend disproportionately more time. For example Edwards et al. (2001) demonstrated that 50% of feral cat location fixes could be found in just 25% of the cat’s home range area. By defining and interrogating these core areas, we can gain better insight into which features (e.g. vegetation, water, landscape features) are preferred by the target animal, allowing us to strategically focus our monitoring and control efforts into these preferred features across the landscape.

An animal’s core area is often defined as the area encompassing the densest 50% of location fixes (Getz et al. 2007). There are many range calculation methods available to analyse spatial data and discriminate these core areas that produce different area estimations for the same animal. For example, Molsher et al. (2005) found that the mean 95% minimum convex polygon (MCP) area was over 33% smaller than the mean 95% fixed kernel density estimate for the same group of feral cats. Several factors contribute to such differences in area estimation, with some methods being more sensitive to different factors than others. For example, it has been shown that excessive bandwidth of the h_REF smoothing parameter overestimates home range size (Naef-Daenzer 1993; Seaman and Powell 1996) in comparison to least-squares cross validation (LSCV) and MCP. Notwithstanding, MCP bias increases with sample size (Burgman and Fox 2003) and is very sensitive to outlying points (Burgman and Fox 2003). By understanding how each of these calculation methods can be influenced by different variables in our data, we can identify which methods more accurately discriminate core areas for the target species, and therefore which method can be used to better inform future research into resource selection and consequently monitoring and control programs.

We used a feral cat GPS collar dataset collected from three geographically distinct areas in Queensland across four years, to estimate home range and core areas using seven commonly used estimation approaches. We compared activity range polygon construction across methods and evaluated each output to identify strengths and limitations for each method. We discuss our findings and make recommendations for the most appropriate method to define core areas to improve monitoring and control of feral cats.

Methods

Study sites

We collected spatial data from GPS collars fitted to feral cats across three geographically distinct national parks in Queensland, Australia (Fig. 1) over four years (2016–2019) as part of a larger feral cat baiting efficacy study (Fancourt et al. 2021, 2022a, 2022b). Detailed site descriptions including location coordinates, bioregion, environ and dominant vegetation communities at each site are provided in Fancourt et al. (2022b).

Fig. 1.

Map showing the location of the three study sites used in the current study within Queensland. Insets on left show (a) Moorrinya National Park, (b) Taunton National Park (Scientific) and (c) Currawinya National Park. Inset at top right shows location of Queensland within Australia.

Taunton National Park (Scientific) (TNP) is located in the subtropical, Brigalow-belt region of central Queensland, approximately 150 km west of Rockhampton (Department of Environment and Resource Management 2011a). During the study period, rainfall at the site varied between years, with above average rainfall of 418 mm received during the five month study period in 2016, but below average rainfall of 88 mm received over the same period in 2017 (Bureau of Meteorology 2020a).

Currawinya National Park (CNP) is located in the semiarid region of south-west Queensland, approximately 25 km north-west of Hungerford. The site encompasses a range of different landscapes ranging from sand dunes and open plains to lakes, saltpans and riparian zones (Queensland Parks and Wildlife Service 2001). The study season was very dry and the site received just 11 mm of rainfall across the five month study period (Bureau of Meteorology 2020b).

Moorrinya National Park (MNP) is located 85 km south of Torrens Creek in the Desert Uplands region of central Queensland (Department of Environment and Resource Management 2011b). During the three-month study period, the study site received a slightly below average rainfall of 22 mm (Bureau of Meteorology 2020c).

Results

Comparison of home range (90% isopleth) and core area (50% isopleth) estimations

Despite the core area containing 50% of fixes, each method estimated a core area that was disproportionally smaller than the corresponding home range estimate (see Fig. 2). R-LoCoH demonstrated the largest proportional core area at 46% of the home range size, whereas the α-hull demonstrated the smallest proportional core area at just 9.5% of the corresponding home range size.

Fig. 2.

Mean proportional differences between home range (90% isopleth) and core area (50% isopleth) estimations for all cats in this study across all seven calculation methods. The white text on bars indicates the percentage of the home range that the core area encompasses. Error bars indicate standard deviation.

Discussion

We compared and contrasted the reliability and suitability of seven home range calculation methods to robustly describe core areas for feral cats. The choice of method had a strong influence on the size and accuracy of the core area (50%) estimation and any comparisons made thereafter. Five of the seven methods were plagued by high variability, high error rates, inconsistency across nonuniform datasets and irregular distributions resulting in inappropriate polygon placement and potential over/under exaggeration of core area estimations. The k-LoCoH and the α-hull appeared superior to the other methods due to their ability to overcome variations in data distribution and sample size, resulting in consistent, sharply defined polygons. However, the generally higher precision of the 50% polygon placement, resulting in lower Type I error inclusion, more definite landscape feature conformity and a uniform parameter value (α) suggested that the α-hull was preferable to k-LoCoH. Our findings suggest that the α-hull method is the most suitable choice for determining high activity areas for feral cats, facilitating more robust habitat analyses. Enhancing the cost effectiveness and efficacy of monitoring and control programs could be achieved by identifying preferred vegetation or landscape features of feral cats and strategically placing monitoring and control devices in these locations.

This study supports previous research indicating that male cats typically exhibit larger home ranges compared to females (Liberg et al. 2000; McGregor et al. 2015; Bengsen et al. 2016). However, there was an isolated exception observed at CNP, where the estimation of one female cat (2018_Cat04, as seen in Table S3) significantly amplified the mean differences of r-LoCoH range size between males and females. This exaggeration in results was unique to the r-LoCoH and highlights this methods’ inherent variability. The generally larger size of female ranges at CNP (most arid site) aligns with landscape productivity assessments reported by Bengsen et al. (2016). Likewise, the generally smaller range estimations at TNP in 2017–2016, reflect higher productivity during 2017, attributed to high rainfall prior to the study period, due to Cyclone Debbie.

Our findings are consistent with other studies (Konecny 1987; Edwards et al. 2001) that cats have multiple core areas, particularly in drier regions where the spatial separation of waterholes drives separation of prey, shelter and breeding females. In our study, male cats at the driest site (CNP) demonstrated the largest MCP, kernel (h_REF) and a-LoCoH core area estimations across all study sites. In contradiction, male cats at CNP also demonstrated the smallest k-LoCoH, α-hull and kernel (LSCV) estimations across all study sites. The sensitivity of the MCP and kernel (h_REF) to satellite data clusters (Burgman and Fox 2003) can lead to greatly exaggerated and inaccurate core area estimations (e.g. MCP – Fig. S1A), so the use of these methods for species exhibiting this type of activity is not advised. The smaller and more refined nature of polygons created by hull methods and the kernel (LSCV) demonstrated greater precision in the differentiation between high and low use areas, even when they were greatly dispersed, demonstrating greater accuracy in core area estimations. The use of fixed time interval programming in collars, causing GPS recordings to occur at the same time each day, may result in misleading or biased data clusters due to the capture of repetitive behaviours associated with the time of day/night. To mitigate this potential bias, a rolling fix interval of 11 or 13 h, instead of 12-h intervals, would ensure that fixes are recorded at different times each day. However, repetitive behaviour patterns can also indicate the utilisation of specific landscape features, which can provide valuable insights for management purposes (Bracis et al. 2018; Campbell et al. 2021; Wilson et al. 2023a).

The influence of subjective input through parameter manipulation was of concern for some methods, where parameter selection greatly affected the occurrence of Type I and II errors. The introduction of these error types primarily stems from three factors: inappropriate polygon placement, parameter influence, and nonuniform datasets or distributions. These error types are often negatively associated, where the avoidance of one can lead to an increase in the other. However, minimising both error types is crucial for accurately depicting high activity areas. Kernel methods, which rely on three parameters, present a considerably higher risk of subjective input and error. Grid intersections in kernel methods can be adjusted by modifying the grid size and resolution (extent), whereas the degree of overlap between kernels at the intersections can be altered by changing the smoothing parameter. The default smoothing parameter in kernel (h_REF) is larger than in kernel (LSCV), resulting in area estimations and polygons that are excessively large for accurately representing the core area and the fine-scale habitat use analysis required (Calenge 2006). Refining all three parameters described above to achieve optimal outcomes for each individual cat is exceptionally time-consuming and increases the number of variables that need to be considered during data analyses. Our results demonstrate that r-LoCoH, which readily portrays high error levels, is somewhat erratic and therefore the least reliable LoCoH method, consistent with the findings of Getz et al. (2007). But Getz et al. (2007) also suggest that a-LoCoH outperforms both r-LoCoH and k-LoCoH due to its heuristic parameter being closer to the optimal value across a wider range of datasets. Because we used real ecological data in this study, we cannot know for certain what the optimal value for any of the LoCoH methods were, and it is difficult to draw conclusions from the data without any level of subjectivity. However, given the optimal LoCoH parameter is defined as the value that minimises the total error for the corresponding method (Getz et al. 2007), it is apparent that the heuristic k value for k-LoCoH is closer to the optimal value than for other LoCoH methods due to substantially fewer Type I and II errors at the sample sizes used in this study. Furthermore, employing a single parameter value derived from a mathematical formula like k-LoCoH reduces the risk of subjective input. In contrast to Bengsen et al. (2012), we did not utilise averaged LoCoH parameter values across all cats due to substantial variations in the dataset between individuals and study sites, leading to average parameters that yielded larger and more frequent errors. Additionally, we did not employ the Minimum Spurious Hole Covering (MSHC) rule (Getz and Wilmers 2004), as this technique can be less effective in landscapes with few inhospitable features, as observed in our study sites, and can therefore introduce a significant level of subjective input. The α-hull avoids subjective input by using a default, uniform parameter (α) across the entire dataset, making it a preferable choice over most other methods examined in this study.

The influence of inconsistency in sample size and data distributions (scattered vs clustered) varied according to method. For instance, kernel methods do not converge on the true extent of a home range as sample size increases (Seaman et al. 1999; Getz and Wilmers 2004). These methods also demonstrate higher error rates under such circumstances (Naef-Daenzer 1993; Worton 1995). Our findings of large sample sizes with clustered distributions (Fig. S4.1A) support this hypothesis. However, it is inappropriate to infer that kernels will perform well at small sample sizes, as the increased likelihood of a scattered distribution is likely to yield large Type II errors (see Figs S7B for kernel (h_REF) and S6B for kernel (LSCV)). In practical scenarios, such as this, we cannot know for certain that the cat did not use areas we consider as Type II errors, and lower sample sizes may be the cause of the scattered distributions, rather than the habits of the cat or the nature of the landscape. But it was evident that low sample size with scattered distributions greatly inflate kernel core area size. Autocorrelated kernel density estimate (aKDE) home ranges have been demonstrated to perform well in these scenarios (Fleming et al. 2015). However, because previous studies on feral cat ranges have predominantly focused on kernel (h_REF) methods (Table S5), and aKDE core ranges in our study were nearly identical to kernel (h_REF) (Supplementary Fig. S10), we excluded aKDE from further analysis. Datasets with high sample size and scattered distributions may allow for reasonable MCP core area estimations (Fig. S5A) but this deteriorates rapidly when the distribution becomes clustered (Fig. S1A) or scattered with low sample sizes (Fig. S7A). The high variability demonstrated by the MCP and both kernel methods do not promote confidence in their use for core range estimations (Table S5).

The high variability in core area estimations observed with r-LoCoH was consistent with findings by Getz et al. (2007), although the source of variability was unconfirmed. The ability of the remaining hull methods to produce consistent core area estimation across multiple data distribution types seen in this study (high sample size and clustered; high sample size and scattered; and low sample size and scattered) further demonstrates their superiority over kernel methods, the MCP and r-LoCoH. In our view, opting for a-LoCoH instead of k-LoCoH or α-hull, as observed in Bengsen et al. (2012) and Recio and Seddon (2013), likely results in exaggerated estimations of core area size. The k-LoCoH and α-hull produced estimations that were smaller and more precise across all distribution types seen in this study. However, the α-hull demonstrates lower Type II error when applied to datasets with high sample size and clustered distributions (Fig. S8D) and demonstrates a finer conformity to landscape features (Fig. S9B). Burgman and Fox (2003) state that the α-hull converges on the true distribution with higher sample sizes, but we found that it performed remarkably well across all data distribution types observed in this study.

Our research suggests that the α-hull method is superior for accurately delineating core areas for feral cats. The α-hull’s strength lies in its high precision of polygon placement, low Type I and II errors, strong conformity to landscape features, repeatability and avoidance of subjective input. On average, α-hull core areas demonstrated the smallest comparative size (i.e. 9.5%, see Fig. 2) to corresponding home ranges of all methods studied here. This generates confidence in estimates of core area placements, which in turn instils confidence in subsequent assessments of habitat preference. If the MSHC rule is not applied to the LoCoH estimations, the default α value by adehabitatHR’s CharHull method (Calenge 2011) is more consistently accurate than the LoCoH heuristic values, introduces less potential bias, and is easier to implement. However, due to the process of polygon construction, the α-hull requires considerably longer computing time than other methods studied here. Nevertheless, the utilisation of the α-hull method provides representative information for subsequent analysis of critical habitat types and landscape features selected by cats in these core areas.

There are additional important considerations that may influence the outcomes of home range estimation. Autocorrelation is believed to influence the estimation of range distributions (Swihart and Slade 1985). To reduce such influence, previous feral cat studies have relied upon subsampling of data or low frequency location fixes (Table S5). Accordingly, we restricted location fixes to twice daily to reduce the effects of autocorrelation on estimates, although some likely remain. Methods that account for autocorrelation may permit a higher frequency of fixes to be included in analyses, although the impact on home range estimation in this study is likely to be minimal (Supplementary Material 10). We concur with Noonan et al. (2019) that developing estimators that account for autocorrelation with the precision of polygon placement of such methods such as LoCoH’s or α-hulls, would be highly beneficial. Additionally, temporal variability may influence the results of this study. Cats may alter their space use over time, likely in response to resource abundance, which may affect the outcomes of such studies.

The effectiveness of current control strategies is dependent upon encounter and interaction rates of feral cats with control tools (Fancourt et al. 2021). Therefore, understanding resource selection and behaviour of cats in identified core areas is crucial for improving encounter and interaction rates. Utilising an appropriate method to accurately describe a species’ high use areas is critical for informing future assessments of resource selection. By discerning core areas in a way that best reflects the true high use areas, critical landscape features or vegetation types that are selected for can be more reliably identified and used to inform monitoring and control strategies. By comparing and contrasting the methodological weaknesses and limitations of the seven most frequently used home range calculation methods, this study helps to identify appropriate home range estimation methods to accurately delineate high use areas.

Supplementary material

Supplementary material is available online.