A machine learning approach to managing game bird introductions

Introduction

The large-scale release of captive-bred species for introduction is a common management practice, especially for species with commercial or recreational value (Bilal, 2022; Larsen et al., 2007; Lever, 2005; Long, 1981; Moulton et al., 2018; Parish & Sotherton, 2007). Many countries established formal acclimatization programs during the 19th and 20th centuries, aiming to enrich local fauna, enhance hunting opportunities, or provide food resources through the introduction of non-native species (Lever, 2005; Long, 1981). Throughout the 20th century, the Foreign Game Investigation Program (FGIP)—a United States (U.S.) government-initiative—focused on identifying, acquiring, and evaluating non-native wildlife species for potential introduction, primarily as game animals (Bohl & Bump, 1970; Bump, 1941; Bump, 1951; Bump, 1963; Moulton et al., 2018; Smith, Cropper Jr & Moulton, 2021). This program aimed to supplement native game populations by introducing species with favorable traits (e.g., hardiness, high reproductive potential, and hunting appeal) into areas where native species were scarce or declining. This process often involved experimental releases, habitat assessments, and extensive field trials. Several avian species, particularly within the family Phasianidae, were evaluated and released as part of this initiative (Banks, 1981; Bohl, 1957a; Bohl, 1957b; Bohl & Bump, 1970; Bump, 1941; Bump, 1963; Bump, 1968; Gullion, 1965).

One of the most notable successes was the introduction of the Chukar Partridge (Alectoris chukar). Chukars are ground-dwelling Eurasian Galliformes that favor high-elevation, arid habitats dominated by talus slopes, sparse grasses, and shrub cover (Alcorn & Richardson, 1951; Barnett, 1952; Bohl, 1957a; Christensen, 1970; Christensen, 2020; Galbreath & Moreland, 1953; Gruychev, Dyakov & Dimitrov, 2014; Harper, Harry & Bailey, 1958; Smith, Cropper Jr & Moulton, 2021). They are among the most widely introduced gamebirds worldwide, with both notable successes and failures. In the conterminous U.S., introductions were attempted in at least 42 states, though self-sustaining populations are now confined to ten western states. Their core distribution lies in the Great Basin, extending into eastern Washington, northern Idaho, western Wyoming and Colorado, the northwestern corner of Arizona, and parts of Montana (Christensen, 1970; Christensen, 2020). Successful introductions also occurred in the Hawaiian Islands, New Zealand, and through translocations within the native range, whereas efforts in Australia, western Europe, and southern Africa were largely unsuccessful (Lever, 2005; Long, 1981).

Although programs like the FGIP helped expand gamebird diversity in some regions, they have been criticized for emphasizing a few successful introductions—such as the Chukar—while overlooking many failed attempts where species did not establish breeding populations (Gullion, 1965; Pierce, 1956). Success is often attributed to factors like large release numbers (e.g., Blackburn, Lockwood & Cassey, 2009; Blackburn, Lockwood & Cassey, 2015; Blackburn et al., 2013; Lockwood, Cassey & Blackburn, 2005), as seen in states like Utah (185,911), Oregon (113,675), and Washington (50,900); however, many of these large propagules commonly involved releases after initial introductions had succeeded (Barnett, 1952; Harper, Harry & Bailey, 1958; Moulton & Cropper, 2015; Moulton & Cropper, 2016). Chukars were not uniformly distributed across these states, often traveling significant distances (i.e., 5–50 km) to reach preferred habitats, even when release sites appeared similar (Bohl, 1957a; Galbreath & Moreland, 1953). Despite large-scale releases, several attempts in states including Minnesota (85,000), Wisconsin (43,013), and New Mexico (31,000) failed to establish populations, raising concerns about the effectiveness of large-scale introductions (Moulton & Cropper, 2015; Moulton & Cropper, 2016; Moulton & Cropper, 2019; Moulton et al., 2018). These inconsistencies highlight that large propagule sizes alone cannot guarantee success, and that environmental variability plays a critical role (Moulton & Cropper, 2015; Moulton & Cropper, 2016; Smith, Cropper Jr & Moulton, 2021; Smyth & Drake, 2022; van der Marel, Waterman & López-Darias, 2021). Even so, traditional game management largely depended on broad, qualitative habitat evaluations and trial-and-error introduction strategies (Christensen, 1970; Gullion, 1965; Lever, 2005; Long, 1981; Pierce, 1956), which lacked precision and often ignored important local factors. This not only led to ineffective outcomes but also raised financial and ethical concerns due to high mortality in unsuitable habitats (Bilal, 2022; Gullion, 1965; Madden, Santilli & Whiteside, 2020). Thus, a key challenge moving forward is identifying locations with ideal environmental conditions that truly drive success while minimizing unnecessary losses.

Recent advances in computing technology and comprehensive data sets, particularly over the last two decades, have enabled the use of more sophisticated statistical and quantitative methods to better identify suitable habitats and assess trade-offs between environmental variables (Beery et al., 2021; Elith et al., 2006; Franklin, 2010; Guisan, Thuiller & Zimmermann, 2017; Guisan et al., 2013; Howard et al., 2014; Peterson et al., 2011; Zimmermann et al., 2010). For example, species distribution models (SDMs), can address these limitations by incorporating detailed data and statistical analysis to predict the potential geographic range of a species based on its known locations and habitat characteristics (Austin, 2002; Elith & Leathwick, 2009; Guisan & Thuiller, 2005; Guisan et al., 2013; Miller, 2010; Norberg et al., 2019; Valavi et al., 2022). SDMs are routinely built using advanced algorithms known as machine learning, a type of artificial intelligence that enables computers to learn from data and improve their performance over time without being explicitly programmed to mechanistically predict (Beery et al., 2021; Crisci, Ghattas & Perera, 2012; Olden, Lawler & Poff, 2008; Pichler & Hartig, 2023; Ramampiandra et al., 2023; Zhang & Li, 2017). These algorithms identify patterns in data, make predictions, or classify information based on known samples (Kuhn & Johnson, 2013).

SDMs are commonly used to assess the impacts of future climate change (e.g., Austin & Van Niel, 2011; Pearson & Dawson, 2003) and to map potential invasion corridors after a species has established (e.g., Barbet-Massin et al., 2018; Gallien et al., 2012; Mainali et al., 2015); however, they are less frequently employed to inform intentional species introductions for game management . To address this gap, Smith, Cropper Jr & Moulton (2021) assessed the predictive power of SDMs in identifying suitable sites for Chukar introductions by examining how site-level environmental factors relate to establishment success across the contiguous U.S., using ensemble models based on data from the species’ native range. While their results underscored the role of local environmental conditions in shaping outcomes, the models primarily predicted broad potential ranges rather than identifying precise introduction sites. Moreover, the authors only considered locations with known occurrences, excluding areas without observations. To improve predictive accuracy and inform future efforts, it is essential to compare successful and failed introduction sites—an approach that can help isolate key environmental drivers of establishment and guide more effective introductions.

In this study, we explore the use of machine learning-based SDMs as a practical tool to guide intentional gamebird introductions. Using the Chukar Partridge as a case study, we analyze historical introduction outcomes in Washington—one of the few events with records documenting both successes and failures across a range of perceived suitable and unsuitable habitats. Our first objective was to use SDMs to predict outcomes in select regions and then apply the models across the rest of the state. Our second goal was to evaluate the ability of these models to generalize beyond the training region by testing predictions in neighboring Oregon. By comparing results across both states, we assess model performance under varied outcomes and refine our understanding of the environmental drivers behind successful introductions. Ultimately, this approach aims to improve the precision and efficiency of future introductions, moving away from traditional trial-and-error strategies toward more informed and cost-effective practices.

Materials & Methods

Study area and context

We simulated the Chukar introduction efforts in Washington State, as documented by Barnett (1952) and Galbreath & Moreland (1953) (Fig. 1). Washington is one of the few states where both successful and unsuccessful Chukar releases have been recorded, in areas perceived as both suitable and unsuitable. The state’s environmental diversity—from its wet western regions to its arid eastern landscapes—offers a valuable gradient for examining the factors that influence establishment success (Alcorn & Richardson, 1951; Barnett, 1952; Galbreath & Moreland, 1953; Gohain, 1959). Accordingly, we used the Washington counties where Chukars were introduced as the foundation for our models.

Between 1938 and 1951, over 5,800 Chukars were released across 24 counties. The most consistent success occurred in the drier, eastern counties, which closely resemble the species’ native range (Barnett, 1952; Galbreath & Moreland, 1953; Lever, 2005; Long, 1981). Other attempts failed, likely due to unsuitable land cover. Releases on the western side of the Cascades failed quickly—largely attributed to the region’s damp climate. In eight counties, mainly located in the southeastern quadrant, outcomes were inconclusive.

Following the eventual success in Washington, and given its similar environmental conditions, Oregon initiated its own widespread Chukar release program (Lever, 2005; Long, 1981). Unlike Washington, Oregon focused exclusively on areas deemed environmentally suitable, releasing approximately 50,000 Chukars between 1951 and 1956. By 1967, the population had grown substantially, with an estimated 1,235,000 individuals distributed across the state.

To evaluate factors influencing Chukar establishment success, we first developed SDMs based on introduction records from Washington State. Our initial focus was on counties with variable introduction outcomes, which served as a testing ground to assess the predictive capacity of habitat suitability modeling under uncertain establishment conditions. Using presence records from counties with confirmed successful introductions as the training dataset, we calibrated the model to identify key environmental conditions associated with establishment success and extrapolated habitat suitability predictions across all of Washington to assess broader spatial patterns.

To replicate the approach used by the Oregon Game Commission—whose site selection strategy was based on perceived habitat similarity—we refined our Washington-based model for application in Oregon. Specifically, models were trained using the entire Washington dataset, incorporating both training and testing records, to maximize the representation of environmental variation within that state. By capitalizing on shared environmental features and ecological analogs between Washington and Oregon, we projected habitat suitability across Oregon using these Washington-trained models alone. This approach enabled us to evaluate the generality of SDMs across state boundaries and to critically assess the ecological soundness of Oregon’s management decisions based on habitat similarity.

Results

Spatial suitability predictions

Suitability maps for Washington, generated by each modeling algorithm, are presented in Fig. 2 and compared against known species occurrence points. Predicted suitability is ranked on a continuous scale from 0 to 1, with higher values indicating greater predicted suitability. Across all models, higher suitability was generally concentrated in the central and eastern regions of the state. Among the models, the Ensemble, RF, and XGBoost exhibited the strongest spatial agreement, producing sharply defined high-suitability areas that closely aligned with observed occurrence data. ANN and GAM also captured the overall distribution of suitable areas but showed less spatial precision. In contrast, KNN and SVM showed lower spatial specificity and produced broader and more diffuse suitability patterns.

Suitability predictions for Oregon were notably less consistent across models (Fig. 3). In general, the western portion of the state was predominantly predicted as low suitability, which aligns with the absence of observed occurrences in that region. The Ensemble, RF, and XGBoost models again demonstrated the most structured and spatially focused predictions, showing strong correspondence with known presence locations. Predictions from ANN and GAM were similar in pattern but tended to overpredict suitable areas beyond known occurrence zones. KNN highlighted areas of high suitability but frequently overgeneralized, suggesting potential overfitting in spatial extrapolation. The SVM model produced the most conservative prediction, identifying only a few isolated zones of high suitability near dense occurrence records.

Discussion

This study highlights the utility of machine learning-based SDMs as a robust framework for informing intentional species introductions in game management. By simulating historical introduction efforts of the Chukar Partridge in Washington and projecting predictions into Oregon, we evaluated the capacity of SDMs to identify environmentally suitable habitats and assess their transferability across state boundaries. Overall, our models achieved strong performance, particularly the ensemble, RF, and XGBoost algorithms, which consistently showed high classification accuracy and spatial concordance with known species occurrences.

Importantly, our results highlight the benefits of ensemble modeling. A common challenge in SDMs is selecting an appropriate algorithm, particularly in the context of hyperparameter tuning and avoiding overfitting (Araújo & New, 2007; Guisan, Thuiller & Zimmermann, 2017; Norberg et al., 2019; Valavi et al., 2022; Zhang & Li, 2017). Ensemble approaches or model averaging offer robust solutions to these challenges (Araújo & New, 2007; Guisan, Thuiller & Zimmermann, 2017; Smith, Cropper Jr & Moulton, 2021; Thuiller et al., 2009). In our study, the ensemble model exhibited the highest spatial specificity, with sharply defined high-suitability zones concentrated near known presence points. This level of spatial precision is especially valuable for applied decision-making, allowing managers to prioritize candidate introduction sites with greater confidence.

Notably, individual algorithms can also be highly effective (Hao et al., 2020; Kaky et al., 2020; Norberg et al., 2019; Valavi et al., 2022). Both the RF and XGBoost models performed exceptionally well, which is not surprising given that these are themselves ensemble-based decision tree models—further reinforcing the strength of ensemble learning approaches. At the same time, non-ensemble methods such as ANN and GAM produced promising predictions, consistent with previous studies supporting their use in ecological modeling (Franklin, 2010; Guisan, Edwards Jr & Hastie, 2002; Lek & Guégan, 1999). In contrast, models such as SVM and KNN, while statistically sound, generated broader and more diffuse suitability patterns. These results may reflect challenges associated with high-dimensional feature spaces and sensitivity to local density variations (i.e., curse of dimensionality; Hastie, Tibshirani & Friedman, 2009; James et al., 2013), making such approaches less suitable for fine-scale ecological modeling.

The ability of the Washington-trained models to generalize to Oregon also provides insights into the ecological transferability of SDMs. Despite differences in data origin, the top-performing models demonstrated strong predictive power when applied to Oregon’s introduction landscape. This suggests that shared environmental gradients can facilitate the successful application of models trained in one region to another, especially when care is taken to ensure ecological and climatic congruence (Araújo & Peterson, 2012; Barbet-Massin et al., 2012; Van Der Wal et al., 2009; Velazco et al., 2024). The fact that Oregon’s large-scale Chukar introductions were concentrated in regions our models predicted as highly suitable offers retrospective support for the state’s management strategy—despite its original reliance on qualitative assessments—and further asserts the critical role of habitat conditions in introduction success.

At the same time, model accuracy declined when applied to counties in Washington with historically variable introduction outcomes. This reduction in performance likely reflects the ecological ambiguity of these transitional zones, which lie near the threshold of environmental suitability. However, these areas are particularly valuable for evaluating model resolution and sensitivity, as they capture subtle environmental gradients that may influence establishment outcomes. Notably, the continued strong performance of several models—such as the ensemble, Random Forest, and XGBoost—even in these uncertain contexts, suggests that species distribution models can still provide meaningful insights under ecologically ambiguous scenarios.

Nonetheless, several limitations warrant consideration. As with all modeling frameworks, it is important to choose covariates that are ecologically informative, but also linearly independent to reduce overfitting (Duan et al., 2014; Hastie, Tibshirani & Friedman, 2009; James et al., 2013; Norberg et al., 2019). Undoubtably, all covariates measured in the study can affect local populations; nonetheless, this information was highly correlated and/or redundant, which could lead to model bias (Araújo & Peterson, 2012; Hastie, Tibshirani & Friedman, 2009; James et al., 2013; Strobl et al., 2007). Here we reduced the dimensionality by reducing multicollinearity; however, it is possible the incorporation of other variables or different selection methods (e.g., different correlation thresholds, clustering, replacement, etc.) could improve model performance (Austin & Van Niel, 2011). Additionally, our environmental predictors—though comprehensive—may still omit key variables related to microhabitat conditions, interspecies competition, or anthropogenic influences (e.g., land use changes). Beyond abiotic factors, studies have shown that behavioral dynamics, local species interactions, and biotic events (e.g., grazing pressure, vegetation shifts, or predator abundance) can strongly influence species distribution and productivity (Gruychev, Dyakov & Dimitrov, 2014; Pearson & Dawson, 2003; Wittmann et al., 2016). Finally, our models rely on historical occurrence records, which may be influenced by observer bias or inconsistent survey effort (Phillips et al., 2009; Zizka et al., 2019).

Similarly, pseudo-absence selection requires careful consideration to avoid mischaracterization of habitat suitability (Barbet-Massin et al., 2012; Phillips et al., 2009). From a methodological standpoint, our inclusion of pseudo-absences derived from both successful and failed introduction counties enhances the ecological realism of our models. This approach enabled direct comparisons between suitable and unsuitable environments, rather than relying solely on presence-only data,which are often spatially biased or clustered (Elith & Leathwick, 2009; Peterson et al., 2011; Van Der Wal et al., 2009). To minimize the risk of incorporating false absences and improve model discrimination, we applied environmental filtering and spatial buffering which previous studies have shown improve model predictions (e.g., Barbet-Massin et al., 2012; Phillips et al., 2009; Senay, Worner & Ikeda, 2013; Van Der Wal et al., 2009; Zbinden et al., 2024). However, some areas classified as unsuitable may have been inaccurately represented. For example, Chukars released from captivity are highly mobile and generally exhibit low site fidelity, complicating efforts to accurately infer habitat associations (Bohl, 1957a; Christensen, 1970; Christensen, 2020; Galbreath & Moreland, 1953; Harper, Harry & Bailey, 1958). Additionally, individuals have been documented dispersing more 50 km from introduction points (Bohl, 1957a; Galbreath & Moreland, 1953; Harper, Harry & Bailey, 1958)—well beyond the buffering thresholds used in our analyses. Such mobility may influence model performance, particularly when predictions are extended to regions of uncertain establishment (Barbet-Massin et al., 2018; Velazco et al., 2024). We therefore recommend that future studies critically evaluate assumptions regarding species movement and site fidelity when selecting pseudo-absences, as these methodological choices can substantially affect model outputs.

Despite these challenges, the modeling framework presented here offers a significant advancement over traditional introduction strategies. By combining high-resolution environmental data, robust machine learning algorithms, and empirical records of both success and failure, this approach allows wildlife managers to make more informed and efficient decisions. As species introductions become increasingly subject to ecological, ethical, and economic scrutiny, the integration of SDMs into environmental planning represents a critical step toward effective management practices.

Conclusion

Our study demonstrates that SDMs, especially with machine learning ensemble methods and biologically relevant pseudo-absence selection, can predict gamebird establishment successfully across diverse landscapes. Additionally, our models provided strong predictions for retrospective evaluation and future management planning. The consistency of model performance across different algorithms reinforces the importance of habitat suitability in introduction success, while also highlighting the need for careful consideration of species movement behavior and site fidelity in SDM construction. These findings highlight the effectiveness of machine learning and ensemble modeling in guiding species introductions, reintroductions, and broader conservation strategies.

Supplemental Information