Hypothetically, an area is at higher risk of disease transmission if it is more likely to experience interactions between feral pig and domestic pigs raised outdoors, as these outdoor based pigs can serve as a conduit for disease spread from wildlife to humans. Locations at greatest risk for increased contact between both swine populations are those regions that contain feral pig suitable habitat located near outdoor-raised domestic pig premises, especially those OPO with relatively low levels of bio-security., , Contact between feral pigs and outdoor-raised pigs in California has been documented, as feral pigs are attracted to agricultural regions for food, water and mates., , There is enormous value in identifying agricultural regions with a higher probability of feral pig contact, because these areas could benefit from targeted cost-effective disease surveillance and risk-mitigation strategies to prevent disease transmission. Predicting suitable habitat for feral pigs in combination with spatially characterizing the distribution of OPO can provide an important tool to ascertain possible high-risk areas of contact at the feral-domestic pig interface and identify future disease spillover areas.
Species distribution modeling methods have been widely used in ecological studies and are becoming popular for use in epidemiological investigations of disease transmission between wildlife and livestock. Maximum Entropy , which is one type of SDM, plastic planters bulk allows usage of presence-only data for the species of interest . In combination with biologically-appropriate covariate factors, MaxEnt is able to spatially predict the probability of suitable habitat for a species for a chosen spatial unit . These two parallel trends of expanding feral pig populations and a resurgence of raising domestic swine outside has important implications for disease transmission, which could negatively impact both public health and California’s agricultural industry. To the best of our knowledge, there are no maps characterizing where suitable feral pig habitat overlaps with domestic pigs raised outdoors at the farm-level in California. The overall objective of this study entailed spatially identifying potential high-risk areas of disease transmission between these two swine populations. This objective was achieved by a three-step process: 1) predicting suitable feral pig habitat in California using MaxEnt; 2) mapping the spatial distribution of OPO in California; and 3) identifying high-risk regions where there is spatial overlap between feral pig suitable habitat and OPOs, as potential disease transmission areas.
MaxEnt is an established SDM method that produces an output prediction map containing estimates of the relative probability of suitable habitat areas for the species of interest within each pixel, using presence-only points and predictor rasters . For feral pig presence data, we obtained feral pig hunting tags from 2012-19 that were cleaned and recorded with GPS coordinates by the California Department of Fish and Wildlife . Hunters in California are voluntarily asked to report feral pig harvest locations by submitting hunting tags to CDFW. Using hunting records for presence-points of feral pigs or wild boars has been used in previous studies. CDFW 2012-19 feral pig hunting tags totaled 5,148 after removing duplicates. Due to the large amount of data points, hunting tags were also manually filtered by year as a way to decrease the abundance of points before running models to reduce sampling bias and increase model stability, as suggested by previous analyses of MaxEnt. Publicly available predictor layers online, including biotic and abiotic , were included in variable selection steps, see Table 2.1. These predictors were chosen based on known feral pig behaviors, habitat and food preferences.For instance, AVGMODIS was the annual maximum green vegetation fraction combined with 12 years of normalized difference vegetation index data and relates to food and shrub cover for feral pigs.
Other variables included elevation, as feral pigs may prefer specific altitudes, and nineteen environmental variables from the WorldClim set of 30 year trend climatic factors. Examples of environmental variables used from the WorldClim site included BIO6, which is the minimum temperature of the coldest month, BIO13 which represents precipitation of the wettest month and BIO15 which is the coefficient of variation for seasonal precipitation.MaxEnt models were built in R Statistical Software version 0.98.110253 ©. The following R packages were used to run MaxEnt: dismo, sp, and raster. MaxEnt settings were chosen based on previously published literature and included using 25 random test points, 15 replicates, 5000 maximum iterations and the 10-percentile training for the threshold rule. , , A regularization multiplier of 1 through 5 was assessed to avoid overfitting and the default 1 was determined to be the optimal setting for the final model. Logistic values for output was used as well as cross validation, which separates presence points into 80% training and 20% testing data , using k-fold sub-sampling to fit a model. The relative contribution of each variable in a MaxEnt model was assessed comparing both percent contribution and permutation of importance, averaged over the number of iterations run and ascertained by jackknife tests. Predictors for the final model were assessed using a backward variable selection approach: variables remained at each step if their percent contribution or permutation importance was approximately 10% or more. The response curves generated within MaxEnt showed the predicted probability of suitable feral pig habitat for each individual variable, changing per each level of the predictor. MaxEnt model performance was assessed using the area under the curve of the receiver operator characteristic , averaged over the number of chosen replicate runs. AUC reflects a model’s prediction ability, on a scale of 0 to 1.00, with 0.50 representing random chance. While AUC is a standard diagnostic method to evaluate MaxEnt models, some authors suggest calibrating the AUC , which removes spatial sorting bias by using point-wise distance sampling. A ssb close to 1 indicates no spatial sorting bias, whereas a ssb close to 0 suggests a large spatialbias, and the need to use AUCc. The final model was chosen based on the highest AUCc, collection pot relative to other models. The feral pig-domestic pig risk map was built by overlapping California OPO locations with the final MaxEnt feral pig suitable habitat raster. Between 2014-2019, a list of California OPOs was compiled through various sources advisors, web-based searches . GPS coordinates for all OPO were identified using Google Earth Pro v7.3.3. Additionally, an online survey that contained an interactive map component was built with Survey 123 v3.6. The survey contained 29 questions that consisted mainly of multiple choice questions, with a few open ended questions about the number of animals raised . The survey included questions regarding biosecurity practices, swine health and feral pig presence. This online survey was announced electronically to swine related groups and organizations or conducted in-person at events, such as agricultural fairs. The survey instrument and protocols were reviewed and exempted by the Institutional Review Board of the University of California-Davis . To build a risk map for California, the final MaxEnt model predicting suitable habitat for feral pigs was overlapped with the location of OPOs to categorize areas at greatest risk for disease transmission, due to contact between these two swine populations, and characterize risk at the farm-level. The underlying assumption presumed that direct or indirect contact between feral pigs and domestic pigs raised outdoors is a risk for disease transmission. The probability of suitable habitat for feral pigs was extracted from the final MaxEnt model for each OPO location, using the Sample Raster Value tool in QGIS and added to the OPO shape file. Then the Kernel Density tool in QGIS was used to make the risk map, matching the 270m x 270m resolution of the MaxEnt model and using the MaxEnt model probabilities as weights.
Additionally, we used a radius of 5 km at each OPO location, which was an extrapolated average estimate from US based studies that measured home range of feral pigs, understanding that home ranges vary depending on age and gender of animal, as well as resource availability. The Kernel Density map was overlaid with the final MaxEnt model. The final MaxEnt model was chosen based on the highest AUCc of 89.7, relative to other models . Probability values that predict suitable habitat were divided into five equal interval categories: minimal ; low ; moderate ; high ; and extremely high , with 0.87 being the highest predicted probability in the final MaxEnt model. Areas with the highest likelihood of suitable feral pig habitat in California included the north coast from Mendocino County all the way south along the coast to Santa Barbara County, and counties that border these coastal counties . Additionally, suitable habitat areas included the foothills of the Sierra mountains, from Shasta County south to Tulare County. Least likely suitable habitat included the Central Valley and eastern counties of California, from the most northern county of Modoc all the way to Imperial County in the south.Five variables were identified as significant in predicting suitable feral pig habitat in the final model based on 2017 hunting tags . The five significant variables were AVGMODIS, Elevation, BIO6, BIO13 and BIO15. All five variables provided approximately 10% or more percent contribution and permutation importance to the final model. . The jackknife test results provided more information regarding the importance of each variable in the final model . For example, BIO15 was the variable with the highest gain when used alone and elevation had the most information that was not available in the other variables. The response curves for the significant five variables indicated the predicted suitability range of each variable for feral pigs . For instance, feral pigs are predicted to prefer vegetative cover of at least 60% or more. The risk map reflects areas at greatest risk for contact between feral swine and domestic pigs raised outdoor and subsequent potential disease transmission . Areas with the most risk for contact between these two swine populations are denoted in orange or red, with sharper colors representing denser clustering of OPO. The counties with the highest likelihood of suitable feral pig habitat and densest clustering of OPO included: Sonoma, Marin, Napa, Yolo, Nevada, Mendocino and Lake counties. Areas at lowest risk include the full eastern edge of California, which includes the Cascadian and Sierra Nevada Mountain ranges as well as deserts in the south. Table 2 categorizes the distribution of OPO at each level of probable suitable feral pig habitat using the final MaxEnt model levels. The results indicated that 49.18% of OPO are located near extremely high or highly suitable feral pig habitat.In this study, we built a feral pig suitable habitat prediction model for California using MaxEnt at a fine scale of 270m x270m. Significant predictors of suitable feral habitat included precipitation, minimum temperature, elevation, and percentage of vegetation. Additionally, this study overlapped MaxEnt predicted suitable feral pig habitat and outdoor-raised pig operations to create a risk map for potential disease transmission in California at the feral pig-domestic pig interface. To the best of our knowledge, this is one of the first studies that identified areas at risk for feral and domestic pig contact in California. Although previous studies discussed the possibility of feral pig populations spreading disease to outdoor-raised pigs at the county level, to our knowledge, this is the first study to predict risk at the farm-level in California. Since the exact location of most feral pig populations is unknown, species distribution predictive methods like MaxEnt are important to understand where feral pigs could potentially interface with domestic swine raised outside, either currently or in the future. Our final MaxEnt prediction model provides a more informative picture of suitable habitat for feral pigs than previous studies, which only showed single presence points or reported feral pigs at the county level, even if only one feral pig was identified in that county. For instance, although previous county-level maps indicated that all California counties except for Imperial County harbored feral pigs, our MaxEnt model shows almost no suitable habitat in an additional five counties: Modoc, Mono, Alpine, Lassen and Inyo. This result may indicate that few feral pigs have been seen in those counties. Additionally, the final MaxEnt model was based on a fine spatial scale and indicated heterogenous suitable habitat, not a uniform distribution, for each county, which is compatible with the fact that feral pigs need shrub cover and food to survive, which would not be found in cities or deserts.