Wildlife Movement & Telemetry

The Range and Distribution Mapping and Analysis Project (RADMAP) in the California Department of Fish and Wildlife's (CDFW) Biogeographic Data Branch (BDB) develops and maintains spatial models for use in conservation decision making, including species range maps and species habitat models. RADMAP is building a library of vetted species range maps and habitat models within California for use by CDFW staff and partners.

This species habitat model (SHM) is a continuous layer depicting a species’ predicted habitat associations within each cell, represented as a value between 0 and 1. Values closer to 1 depict a higher relative probability of habitat use within that cell. The focal taxon may or may not actually occur in areas predicted with a high relative probability of habitat use; habitat may be suitable but unoccupied, particularly for taxa with small and/or declining populations or limited mobility. Models may not accurately reflect all habitats used per taxon due to a dearth of presence data, thereby limiting the scope of environmental space represented by the SHM. Users should refer to the validation metrics and consider the level of uncertainty associated with the model when interpreting model outputs. Areas with high relative predicted use values may be prioritized for locating survey or monitoring sites for scientific studies aiming to conserve and protect focal taxa.

Occurrence records were obtained from a large, collated dataset as indicated below. To reduce sampling bias and avoid model overfitting, which can reduce model applicability to unsampled areas, we excluded spatially autocorrelated presence records. Owing to the breadth of our study region, we accounted for topographical heterogeneity using a digital elevation model and filtered occurrence records based on this new raster at three distinct distances. Areas with species occurrences and high topographical heterogeneity were filtered at closer Euclidean distances. For potentially ecologically relevant environmental covariate inputs we computed a Pearson correlation coefficient matrix to assess the strength of association among variables; those that were highly autocorrelated (greater than or equal to 0.7) were removed from further analyses. All covariates were continuous and formatted at a resolution of 30 m.

Potential habitat use was estimated using a maximum entropy approach (Maxent) implemented via R language that relied on presence data and comparisons between environmental covariate values at presence localities and those at randomly selected background sites (Phillips et al. 2006). To demarcate the specific geographical area used for model calibration, background locations were selected via local adaptive convex-hull polygons based on known species’ dispersal and movement limitations. This allowed RADMAP to exclude uncolonized suitable habitat, potentially due to dispersal barriers or inhibitory biotic interactions; it also precluded overfitting models to environmental conditions immediately adjacent to occurrence records, thus improving predictive performance. For each model, we randomly sampled > 10,000 background locations. Five feature-class combinations and five regularization-multiplier settings were adjusted rather than using default Maxent settings to improve model fit. We considered a range of regularization multipliers in integer-sized increments from 1 to 5, then divided data into training and test groups using geographically structured k-fold cross validation (k = 4) to reduce overfitting to environmental conditions among spatial partitions. A total of 25 models were run during this phase of model development. Additionally, 25 full models were run using all available presence data.

Models were evaluated with multiple statistics, including test omission error rate (OER) and true skill statistic (TSS), which are threshold-dependent metrics. Second, we generated receiver-operating-characteristic curves and assessed model performance using area-under-the-curve (AUC) analyses for test data, a threshold-independent metric. AUC calculates an average value for the k-folds used on the analysis and assesses the difference between AUC training and test data (AUCdiff), the latter of which is used to quantify overfitting (i.e., lower values indicate better fits). The training model with the lowest OER and highest AUC and TSS values for both test and training models was considered the top model. We calculated the contribution percent of each predictor variable to the top model to identify the explicit role of each in influencing the distribution of a species. Models were extrapolated to the taxon’s range to predict across all potentially occupied areas within California.

Results of the Maxent model are included in the pdf attachment, including top model review score, validation metrics, model output details, covariate response curves, percent contribution of covariates to the top model, and a full list of covariates included in the model.

The top model results are linked here: https://nrm.dfg.ca.gov/FileHandler.ashx?DocumentID=246780.

The following data sources were collated for the purposes of range mapping and species habitat modeling by RADMAP. Each focal taxon’s location data was extracted (when applicable) from the following list of sources. BIOS datasets are bracketed with their “ds” numbers and can be located on CDFW’s BIOS viewer: https://wildlife.ca.gov/Data/BIOS.

• California Natural Diversity Database,

• Terrestrial Species Monitoring [ds2826],

• North American Bat Monitoring Data Portal,

• VertNet,

• Breeding Bird Survey,

• Wildlife Insights,

• eBird,

• iNaturalist,

• other available CDFW or partner data.

Please refer to the Range Map and Species Habitat Model Use Case Guidance document on how best to interpret RADMAP outputs, including range maps, continuous SHMs, and categorical SHMs. Specifically, users should follow these guidelines to determine which products to utilize for conservation, management, and policy decision making use cases: https://nrm.dfg.ca.gov/FileHandler.ashx?DocumentID=222269.

The Range and Distribution Mapping and Analysis Project (RADMAP) in the California Department of Fish and Wildlife's (CDFW) Biogeographic Data Branch (BDB) develops and maintains spatial models for use in conservation decision making, including species range maps and species habitat models. RADMAP is building a library of vetted species range maps and habitat models within California for use by CDFW staff and partners.

The categorical species habitat model (SHM) is derived from a continuous SHM, splitting the continuous model output into predicted habitat and non-habitat. This map simplifies continuous SHMs, pinpointing areas within a species’ predicted habitat with a high, medium, and low relative probability of habitat use. Habitat (values higher than the minimum training presence threshold) is divided into low, medium, and high categories based on expert-chosen predicted thresholds. The focal taxon may or may not actually occur in areas predicted with a high relative probability of habitat use; habitat may be suitable but unoccupied, particularly for taxa with small and/or declining populations or limited mobility. Models may not accurately reflect all habitats used per taxon due to a dearth of presence data, thereby limiting the scope of environmental space represented by the SHM. Users should refer to the validation metrics and consider the level of uncertainty associated with the model when interpreting model outputs. Areas with high relative predicted use values may be prioritized for locating survey or monitoring sites for scientific studies aiming to conserve and protect focal taxa.

Occurrence records were obtained from a large, collated dataset as indicated below. To reduce sampling bias and avoid model overfitting, which can reduce model applicability to unsampled areas, we excluded spatially autocorrelated presence records. Owing to the breadth of our study region, we accounted for topographical heterogeneity using a digital elevation model and filtered occurrence records based on this new raster at three distinct distances. Areas with species occurrences and high topographical heterogeneity were filtered at closer Euclidean distances. For potentially ecologically relevant environmental covariate inputs we computed a Pearson correlation coefficient matrix to assess the strength of association among variables; those that were highly autocorrelated (greater than or equal to 0.7) were removed from further analyses. All covariates were continuous and formatted at a resolution of 30 m.

Potential habitat use was estimated using a maximum entropy approach (Maxent) implemented via R language that relied on presence data and comparisons between environmental covariate values at presence localities and those at randomly selected background sites (Phillips et al. 2006). To demarcate the specific geographical area used for model calibration, background locations were selected via local adaptive convex-hull polygons based on known species’ dispersal and movement limitations. This allowed RADMAP to exclude uncolonized suitable habitat, potentially due to dispersal barriers or inhibitory biotic interactions; it also precluded overfitting models to environmental conditions immediately adjacent to occurrence records, thus improving predictive performance. For each model, we randomly sampled > 10,000 background locations. Five feature-class combinations and five regularization-multiplier settings were adjusted rather than using default Maxent settings to improve model fit. We considered a range of regularization multipliers in integer-sized increments from 1 to 5, then divided data into training and test groups using geographically structured k-fold cross validation (k = 4) to reduce overfitting to environmental conditions among spatial partitions. A total of 25 models were run during this phase of model development. Additionally, 25 full models were run using all available presence data.

Models were evaluated with multiple statistics, including test omission error rate (OER) and true skill statistic (TSS), which are threshold-dependent metrics. Second, we generated receiver-operating-characteristic curves and assessed model performance using area-under-the-curve (AUC) analyses for test data, a threshold-independent metric. AUC calculates an average value for the k-folds used on the analysis and assesses the difference between AUC training and test data (AUCdiff), the latter of which is used to quantify overfitting (i.e., lower values indicate better fits). The training model with the lowest OER and highest AUC and TSS values for both test and training models was considered the top model. We calculated the contribution percent of each predictor variable to the top model to identify the explicit role of each in influencing the distribution of a species. Models were extrapolated to the taxon’s range to predict across all potentially occupied areas within California.

Results of the Maxent model are included in the pdf attachment, including top model review score, validation metrics, model output details, covariate response curves, percent contribution of covariates to the top model, and a full list of covariates included in the model.

The top model results are linked here: https://nrm.dfg.ca.gov/FileHandler.ashx?DocumentID=246780.

The following data sources were collated for the purposes of range mapping and species habitat modeling by RADMAP. Each focal taxon’s location data was extracted (when applicable) from the following list of sources. BIOS datasets are bracketed with their “ds” numbers and can be located on CDFW’s BIOS viewer: https://wildlife.ca.gov/Data/BIOS.

• California Natural Diversity Database,

• Terrestrial Species Monitoring [ds2826],

• North American Bat Monitoring Data Portal,

• VertNet,

• Breeding Bird Survey,

• Wildlife Insights,

• eBird,

• iNaturalist,

• other available CDFW or partner data.

Please refer to the Range Map and Species Habitat Model Use Case Guidance document on how best to interpret RADMAP outputs, including range maps, continuous SHMs, and categorical SHMs. Specifically, users should follow these guidelines to determine which products to utilize for conservation, management, and policy decision making use cases: https://nrm.dfg.ca.gov/FileHandler.ashx?DocumentID=222269.

The project lead for the collection of this data was Bob Stafford. Elk (4 adult females, 2 adult males) from the Chimineas herd were captured and equipped with Lotek GPS collars (GPS3300L, Lotek Wireless, Newmarket, Ontario, Canada) and ATS GPS collars (G2000, Advanced Telemetry Systems, Isanti, Minnesota, USA), transmitting data from 2005-2008. The study area was within the La Panza Elk Management Unit, north adjacent to State Highway 166 and largely residing within Carrizo Plains Ecological Reserve with some movement into Los Padres National Forest. State Highway 166 is a likely barrier to movement as no collared elk moved southward across the road to additional habitats. The Chimineas herd contains short distance, elevation-based movements likely due to seasonal habitat conditions, but this herd does not migrate between traditional summer and winter seasonal ranges. Instead, the herd displays a residential pattern, slowly moving up or down elevational gradients. Therefore, annual home ranges were modeled using year-round data to demarcate high use areas in lieu of modeling the specific winter ranges commonly seen in other ungulate analyses in California. GPS locations were fixed at 1-13 hour intervals in the dataset. To improve the quality of the data set as per Bjørneraas et al. (2010), the GPS data were filtered prior to analysis to remove locations which were: i) further from either the previous point or subsequent point than an individual elk is able to travel in the elapsed time, ii) forming spikes in the movement trajectory based on outgoing and incoming speeds and turning angles sharper than a predefined threshold , or iii) fixed in 2D space and visually assessed as a bad fix by the analyst.

The methodology used for this analysis allowed for the mapping of the herd’s annual range based on a small sample. Brownian Bridge Movement Models (BBMMs; Sawyer et al. 2009) were constructed with GPS collar data from 5 elk in total, including 7 year-long sequences, location, date, time, and average location error as inputs in Migration Mapper to assess annual range. Annual range BBMMs were produced at a spatial resolution of 50 m using a sequential fix interval of less than 27 hours and a fixed motion variance of 1000. Population-level annual range designations for this herd may expand with a larger sample, filling in some of the gaps between high-use annual range polygons in the map. Annual range is visualized as the 50th percentile contour (high use) and the 99th percentile contour of the year-round utilization distribution.

The project lead for the collection of this data was Bob Stafford. Elk (7 adult females, 3 adult males) from the Cedar Canyon herd were captured and equipped with Lotek GPS collars (LifeCycle 800 GlobalStar/ GPS3300L, Lotek Wireless, Newmarket, Ontario, Canada) and ATS GPS collars (G2000, Advanced Telemetry Systems, Isanti, Minnesota, USA), transmitting data from 2005-2006 and 2015-2017. The study area was within the La Panza Elk Management Unit, north of State Highway 58 in the Temblor Range and largely residing on private lands. The Cedar Canyon herd contains short distance, elevation-based movements likely due to seasonal habitat conditions, but this herd does not migrate between traditional summer and winter seasonal ranges. Instead, the herd displays a residential pattern, slowly moving up or down elevational gradients. Therefore, annual home ranges were modeled using year-round data to demarcate high use areas in lieu of modeling the specific winter ranges commonly seen in other ungulate analyses in California. GPS locations were fixed at 1-13 hour intervals in the dataset. To improve the quality of the data set as per Bjørneraas et al. (2010), the GPS data were filtered prior to analysis to remove locations which were: i) further from either the previous point or subsequent point than an individual elk is able to travel in the elapsed time, ii) forming spikes in the movement trajectory based on outgoing and incoming speeds and turning angles sharper than a predefined threshold , or iii) fixed in 2D space and visually assessed as a bad fix by the analyst.

The methodology used for this analysis allowed for the mapping of the herd’s annual range based on a small sample. Brownian Bridge Movement Models (BBMMs; Sawyer et al. 2009) were constructed with GPS collar data from 9 elk in total, including 17 year-long sequences, location, date, time, and average location error as inputs in Migration Mapper to assess annual range. Annual range BBMMs were produced at a spatial resolution of 50 m using a sequential fix interval of less than 27 hours and a fixed motion variance of 1000. Population-level annual range designations for this herd may expand with a larger sample, filling in some of the gaps between high-use annual range polygons in the map. Annual range is visualized as the 50th percentile contour (high use) and the 99th percentile contour of the year-round utilization distribution.