Introduction

African swine fever (ASF) is a swine disease caused by a large double-stranded DNA virus. The ASF virus (ASFV) is a member of the Asfarviridae family and is the sole member of the Asfivirus genus1. ASFV can be transmitted by direct contact between pigs, consumption of contaminated food, blood, pork products, and soft tick vectors2. One significant factor contributing to its widespread dissemination is its high environmental stability. The virus can survive for extended periods in various materials, such as soil, water3, and contaminated equipment4. ASF has spread extensively, causing economic losses in agriculture and animal husbandry5 and affecting the wild boar population in natural habitats across Africa, Europe, the Americas, and Asia6,7. Wild boars serve as reservoirs for ASF in some countries, highlighting the importance of addressing preventing disease transmission between wild boars and domestic pigs8,9. Outbreaks of ASF in wild boar populations have been detected in areas near forested regions of Lao PDR, and Vietnam. Contact, both direct and indirect, between domestic and wild pigs was observed in these communities. Researchers have observed seasonal patterns in wild boar sightings. Discarded pig carcasses have been reported along roads and at forest edges, where they are susceptible to animal scavenging, and wild boars, posing a risk of viral transmission10. ASF infection in a wild boar (Sus scrofa) carcass was reported in the northwestern part of Singapore’s main island on February 5, 2023. Additionally, Dermacentor auratus ticks collected from the area tested positive for ASF11. While S. scrofa is a common wild pig species in Southeast Asia12, Sus barbatus (Bearded pig) was reported to be infected with ASF in Malaysia in 202113. Similarly, Sus philippensis (Philippine warty pigs) tested positive for ASF based on carcasses and meat samples collected from wild pig deaths in Abra province, Philippines, in May 202114.

In contrast, Bush pigs (Potamochaerus spp.) and warthogs (Phacochaerus spp.) found in Africa are natural reservoirs for African Swine Fever15. Wild boar populations in other regions have also been identified as a vital animal reservoir for the ASFV, particularly in Europe. Wild boar can host and spread ASFV, even when not displaying clinical signs16. ASFV transmission cycles have adapted from the original association between soft ticks and warthogs in Africa. The growing human population and increasing domestic pig populations are drivers of ASF spread to previously unaffected areas, resulting in transmission cycles between domestic pigs and wild boars in their habitats17,18.

In Thailand, domesticated wild boars are raised on farms and in backyards19, while free-ranging wild boars have been reported in some protected areas20. However, a study of the distribution and estimated number of free-ranging wild boar across the country has not been conducted to date. This study aimed to predict the geographical distribution of free-ranging wild boar in Thailand by applying a spatial modeling approach and to estimate the number of free-ranging wild boar at both grid and administrative unit levels. The study also provided information on habitat separation between domestic pigs and free-ranging wild boars to prevent ASFV spillover, particularly in areas near forests or along Thailand’s borders with countries that have reported ASF outbreaks. Findings can support the strategic planning and risk mitigation for ASF infection in free-ranging wild boars and domestic pigs while providing information for the surveillance of ASF in Thailand and Southeast Asia.

Results

Geographical distribution of wild boars in Thailand

The data on wild boar occurrences collected through the SMART patrol system managed by the DNP showed that there were 957 sightings of live wild boars, 64,337 instances of footprints, and 3,943 instances of dung observed in 2021 throughout the country. Figure 1 illustrates that the predictive maps indicated the population of wild boars was distributed throughout the country, with higher concentrations in the northern and western regions of the country, particularly in the border areas.

Fig. 1
figure 1

Predicted wild boar distribution maps in Thailand: A map modeled using count data (A), a map modeled using binary data (B), and a map modeled by combining the two models (C).

Full size image

Regarding the approach using a binary RF, forest was the environmental factor most influential in the distribution of wild boars in Thailand, followed by human population density, elevation, rainfed cropland, distance to water bodies, and irrigated cropland, respectively (Fig. 2a). The relationship between the fitted function and the predictor variables, modeled using a binary RF with 10 bootstrap samples, is shown in Fig. 2b–2g. Three environmental factors, including forest cover, elevation, and distance to water bodies, had a positive association with the predicted values, whereas human population density, rainfed, and irrigated croplands showed a negative association.

Fig. 2
figure 2

Key variables and partial plots of the fitted function modeled using binary random forest: Important variables (a) and partial plots of the fitted function for forest cover (b), human population density (c), elevation (d), rainfed cropland (e), distance to water bodies (f), and irrigated cropland (g).

Full size image

All count and binary models demonstrated high predictive power, while the evaluation of final outputs by measuring the GOF between the observed and predicted values indicated moderate accuracy (Table 1). For the count models, the correlation coefficient was 0.92, and the RMSE was 0.97. The binary models showed a high predictive power with an AUC ranging between 0.994 and 0.998 for the training sets and ranging from 0.967 to 0.989 for the test sets (ROC plot shown in the supplementary information). Regarding the final output evaluation, the correlation ranged from 0.568 to 0.614, and the RMSE was between 1.32 and 1.41.

Table 1 Evaluation of predictive performance using random forest models with count data, binary data, and combined outputs.
Full size table

Estimated wild boar population in Thailand

The number of wild boars derived from the 10 data sets was compiled and mapped onto administrative boundary maps at the sub-district, district, and provincial levels. Table 2 displays the estimated total number of wild boars for each set. The average number of wild boars in Thailand was estimated at 307,061 individuals (SD = 4,672).

Table 2 Estimated wild Boar population using 10 bootstrap samples.
Full size table

The estimated number of wild boars in each province is shown in Table 3. The highest number was found in Chiangmai (31,413), followed by Mae Hong Son (21,551), Tak (20,912), Kanchanaburi (20,873), and Nan (14,307).

Table 3 Estimated wild boar population in Thailand, ranked by Province.
Full size table

Interface areas between wild boars and domestic pig farms in Thailand

The interface area between wild boars and domestic pig farms is shown in Fig. 3. Sub-districts with a high number of wild boar (more than 500 individuals) and a high number of domestic pig farms (more than 10 farms per sub-district) were identified as high-risk interface areas.

Fig. 3
figure 3

Analysis of interface areas between high-density wild boar populations and domestic pig farms: A map of wild boar distribution at the sub-district level classified by the predicted population (A), a map of domestic pig farms at the sub-district level (B), and a map of interface areas between high-density wild boar populations (> 500 heads per sub-district) and high concentration of domestic pig farms (> 10 farms/subdistrict) (C).

Full size image

As shown in Table 4, the study covered 55 sub-districts across 29 districts in 8 provinces, indicating high interface between wild boar habitats and domestic pig farm locations. Most of the selected sub-districts were located in the north-west of the country.

Table 4 High interface areas between estimated wild boar populations and domestic pig farms in Thailand.
Full size table

Discussion

African swine fever (ASF) in wild boars (Sus scrofa) poses a significant challenge in Southeast Asia. The transmission of ASF from domestic pigs to wild boars in Southeast Asia highlights the importance of preventing disease transmission in the interface between humans, domestic animals, and wildlife10. Areas with high likelihood of wild boar presence are concentrated in Thailand’s western and northern regions, indicating the fertility of these regions. This suggests that surveillance for ASF in wild boars along the Thailand-Myanmar border is crucial.

In previous studies, wild boar distribution was estimated for managing ASF and conservation purposes using various techniques, such as a generalized linear model21, habitat quality assessment22, logistic regression, and multi-model inference23, maximum entropy (Maxent) models24. Although multiple methods are available for predicting wild boar habitat, we utilized the Random Forest (RF) algorithm to forecast wild boar distribution and abundance across Thailand. This technique stands out for its high accuracy, robustness, ability to identify crucial features, adaptability, and scalability. Its effectiveness lies in addressing overfitting by aggregating multiple decision trees, thereby reducing susceptibility to noise and outliers within the dataset25. This approach not only streamlines the prediction process but also minimizes the time and costs associated with relying solely on expert opinions. Our model is specifically tailored for Southeast Asian countries, aiming to predict wild boar habitats in these regions for enhanced ASF surveillance. This technique was a pilot project that can be easily applied across the region, and we have compiled data sources containing essential geographic and demographic information accessible to each country. Examples include GlobCover for land cover data, water bodies, forests, and both rain-fed and irrigated croplands, elevation data, and human population data. The choice of these factors was based on findings from previous studies that identified them as influential in wild boar habitat determination3,22,23, serving as the foundation for our model.

Our findings emphasize the significance of forested areas and elevation as pivotal factors in detecting wild boar populations. Those were consistent with the previous studies, in which these animals were more prevalent in forested areas3,22 and highlands26. The relationship between the distance to water bodies and the detection of wild boars revealed an initially positive association that decreases beyond a certain point. Wild boars tended to avoid areas too close to water bodies, perhaps due to disturbances from humans or predators27. However, detection decreases when the distance becomes too far; it might limit access to water and reduce their presence. The graph in (Fig. 2f) indicates an optimal distance where wild boars balance, minimizing disturbances while maintaining easy access to water resources. It was different from the previous studies in that wild boar could be detected to decrease when the distance to water increased23. The moderate level of rained cropland was the most incredible regarding wild boar density, consistent with the results of the agriculture area in the previous study in which a relatively high level of rained cropland led to a low possibility of detecting the wild boar21. Additionally, irrigated cropland built by humans and human population density had negative associations in detecting wild boar populations, as shown by our fitted function model (Fig. 2). This might explain why wild boars tend to avoid areas with high human disturbances27. Additional variables may include identifying habitat associations of wild boars, subject to data availability, including seasonal patterns and carnivore diversity21.

The count and binary models showed more accurate predictions than the combination model. This indicates that the count and binary models are suitable for RF wild boar habitat prediction. We estimated the wild boar population using the Smart Patrol System, unlike other techniques that used previously recorded wild boar density data21. However, population density can be calculated based on our estimation method using the wild boar count for each grid area as required. The RF estimation of wild boar population showed that there was a slight variation of approximately 13% in the number of wild boars within a 10-kilometer radius and significant interface areas between wild boar populations and domestic pig farms were identified in the northwest of Thailand, including Chiangmai, Mae Hong Son, and Tak provinces, while Nan province, on the opposite side of Thailand, borders Laos PDR (Fig. 3; Table 4). It emphasizes that the ASF is not only a potential source of transboundary diseases in the Thailand borders but also poses a risk of disease transmission between domestic pigs and wild boars, particularly in areas with backyard pigs, small-holder pig farms, and commercial farms10. Furthermore, high interface areas between wild boar habitats and domestic pig farms are found in north-central Thailand, particularly in Nakhon Sawan, Kamphaeng Phet, Phitsanulok, and the Southern region, especially Nakhon Si Thammarat. Those results are related to the numerous domestic pig farms with wild boar sightings.

In Thailand, measures to prevent and control African Swine Fever (ASF) include effective monitoring systems for both domestic pigs and wild boars. This surveillance enables the quick detection of outbreaks, allowing for prompt action. Collaborative efforts with neighboring countries, particularly Laos PDR and Vietnam, focus on preventing and controlling ASF in border regions. Risk assessments have been conducted on pig farms to evaluate the risk of ASF transmission28. Biosecurity measures have been implemented, such as quarantining and controlling the movement of animals, decontamination of infected animals, and managing vectors. Farmers are also encouraged to prevent direct contact between domestic pigs and wild boars by taking proactive steps, such as building fences around their farms29.

Additionally, the abundance of wild boar populations indicates biodiversity richness in the area, potentially influencing the presence of other species, such as tigers24. Notably, the frequency of wild boar population surveys may vary in certain regions, depending on the monitoring activities conducted by wildlife rangers using the Smart Patrol System. Although this presents a limitation, our RF model focuses on binary data indicating the presence or absence of wild boar in these areas, demonstrating the model’s relatively high performance.

ASFV infection rates in wild boars in Southeast Asia may be underestimated. The accuracy of estimates depends on each country’s surveillance system, especially data from the Smart Patrol System. When there is a lack of information from the Smart Patrol System, we turn to alternative sources. Data on wild boar populations and ASF detection from the National Wildlife Health System (piloted in Cambodia, Lao PDR, and Vietnam), involving protected area rangers, wildlife rescue centers, community members, livestock and human health personnel, and laboratories, can be integrated into our model10. This collaborative surveillance initiative provides a valuable supplementary data stream for refining and enhancing the accuracy of our predictions. However, other species of wild boars, besides Sus scrofa, are found on islands such as the Philippines and Indonesia; a dataset covering different geographic areas may be necessary to estimate wild boar populations.

Although we used a training set to develop the models and a test set to evaluate them, camera traps placed in areas predicted to have high wild boar abundance will confirm the presence of wild boars in areas inhabited by local communities and highlight the likelihood of encounters between wild boars and humans. Our study focused on wild boar information from non-protected areas, where concerns arise regarding interactions between humans, domestic animals, and wildlife in local communities.

One limitation of our data was that the officer conducting the information collection on the Smart Patrol System had difficulty identifying individual wild boars and confirming the uniqueness of each animal. Additionally, the frequency and duration of wild boar surveys on the trails varied by team and did not follow a consistent pattern across the entire country. It might lead to duplicate counting of the number of wild boars. We then used the observed data of live wild boars, footprints, or dung as evidence of the presence of wild boars in those grids (1 km resolution) to model the probability of wild boar distribution using a binary RF and only the observed data (count data) of live wild boars to model the number of wild boars using a quantitative RF. Therefore, the possibility of duplicate counts was relatively low. However, this publication is the first in Southeast Asia to estimate the number of wild boars. Therefore, a well-planned approach to data collection is essential for obtaining more accurate information for future studies.

We expect that our data and techniques would also be relevant and useful to other Southeast Asian countries for conducting comparable studies. Observational data and datasets from the Smart Patrol System of the Department of National Parks, Wildlife and Plant Conservation (DNP) will help validate predictions made using spatial analysis and modeling techniques to estimate the distribution of free-ranging wild boars. The study will also provide information on habitat overlap between domestic pigs and wild boars to mitigate the risk of ASFV transmission, particularly in areas adjacent to forests or bordering countries that have reported ASF outbreaks. Findings can support the strategic planning and risk mitigation for ASF infection in free-ranging wild boars and domestic pigs while providing information for the surveillance of ASF in Thailand and Southeast Asia.

Materials and methods

We used the data on wild boar occurrences in 2021 collected by the Smart Patrol System of the Department of National Parks, Wildlife and Plant Conservation (DNP) to model the distribution and estimate the population of wild boars in the country. The officers under the DNP have been trained to conduct wildlife surveys in the protected area and to enter data into the database known as “SMART (Spatial Monitoring and Reporting Tool)”23. The collected data on wild boar included the locations of live wild boar, footprints, and dung (Fig. 4).

Fig. 4
figure 4

Spatial distribution of wild boar occurrences in Thailand (2021) Distribution of live wild boar (A), footprint (B), and dung (C).

Full size image

We used a random forest (RF) model to quantify the relationship between wild boar presence and the ecological factors. To model complicated interactions among the predictors, RF has high performance compared with other approaches30. It has been increasingly used in modeling population distribution with highly accurate results31,32,33,34,35. RF is a tree-based method, in which the regression algorithm is started by drawing n bootstrap sub-samples from the original data, followed by growing un-pruned regression trees by randomly sampling m variables from a list of predictors, choosing the best spit from those predictors for each of the bootstrap samples, and building a final predicted value by averaging the predictions of the n trees36. The variable importance is presented by counting the number of times each variable is selected in the different trees, so it is an absolute measure where variables’ importance is assessed according to their relative contribution37.

The predictors used for wild boar distribution prediction (Fig. 5) included environmental data (elevation, distance to water bodies, forested areas, rainfed and irrigated croplands) and demographic data (human population density). Geospatial data, including shapefiles of water bodies and forested areas, were provided by the Land Development Department38. 300-m resolution maps of rainfed and irrigated croplands were extracted from the land cover map provided by the GlobCover project39. A raster map of human population density with a 100-meter resolution was acquired from the WorldPop project40. All predictor layers were resampled to a 1000-meter resolution.

Fig. 5
figure 5

Spatial datasets used as predictor variables for modeling wild boar distribution in Thailand.

Full size image

The variables included (a) elevation (log10 of elevation (meters)), (b) distance to water bodies (meters), (c) human population density (log10 of population density (people per square kilometer)), (d) forest (proportion per square kilometer), (e) rainfed cropland (proportion per square kilometer), and (f) irrigated cropland (proportion per square kilometer).

We applied a two-component approach to the model: count data (number of wild boar observed) and binary data (presence or absence of wild boar)41,42. For modeling the presence/absence of wild boar, grids with evidence of wild boar presence, including live sightings, footprints, or dung were assigned a value of 1 (presence), while the rest were assigned a value of 0 (absence) and modeled using binary RF. To model wild boar population density, areas where wild boar were observed were selected and analyzed using a quantitative RF model. Due to the absences being distributed within a given distance of the presence sites, we decided to bootstrap using various distance values. Ten sets of absence data were randomly selected from 1 to 10 km from the presence sites, encompassing the home range and habitat selection of wild boar43. If the selected range for the absence points is too far from the observed animal locations, it may inaccurately characterize those areas as “unsuitable” for their natural habitat. This could lead to the model estimating a larger area for detecting wild boar around the observed locations, consequently inflating the estimated number of wild boars in those regions. Two times the number of positives was randomly selected for each bootstrap. Then, each set was randomly divided into two equal parts: a training set was used to train the model, and a test set was used to evaluate the models. To assess the predictive performance, two statistical metrics, namely the correlation coefficient (COR) and root mean square error (RMSE), were used for quantitative RF models, and the area under the receiver operating characteristic curve (AUC)44 was used to evaluate the predictive power of the binary RF models. For the final step of modeling, we combined the predictive values of both approaches to produce a final map of predictive values. We evaluated the predictive power of the final models using two statistical metrics: the correlation coefficient (COR) and the root mean square error (RMSE). The packages “randomForest”30 and “hydroGOF”45 available in R (R Core Team, Vienna, Austria) were used for the RF model and goodness-of-fit assessment, respectively. The parameters set for RF models included mtry = 5, ntree = 500, nodesize = 5.

We identified areas with high interaction between predicted wild boar populations and locations of domestic pig farms. Firstly, the distribution map of wild boar at the pixel level obtained from the previous step was aggregated to the sub-district level and classified into five categories. Secondly, the data on domestic pig farms surveyed annually in Thailand in 2021 by the Department of Livestock Development (DLD) were aggregated at the sub-district level and classified into five categories. Finally, sub-districts with high densities of both estimated wild boar populations and domestic pig farms were selected.