Species distribution modeling in regions of high need and limited data: waterfowl of China

The goal of this work was to develop high resolution spatial distribution maps for waterfowl species across seasons, providing a critical model input that would pave the way for informed management in a data-restricted region. Although there was a scientific deficit for producing data-driven models using newer SDM techniques (Guisan et al. 2013), the societal need to understand where these birds might occur both individually and cumulatively across species remained. Using the best data available, we took a conventional approach of mapping habitat suitability and combined it with model control measures (range masking) to develop and validate habitat models for each species and combine these into species richness data layers. Such data layers, or any of the species specific layers, can now be directly incorporated as inputs into disease models (Prosser et al. 2016), critical habitat identification modeling, or any other need that requires estimation of a species most probable use of space, thereby resolving the challenge of incorporating information concerning wild birds into such efforts.

The key finding of this work was the seasonal changes in species distribution throughout the annual cycle. While there was variation between species, wintering distributions were predominately concentrated in the warmer and lower elevation regions of the southeast while the breeding distributions were more evenly spaced across the northern latitudes and higher elevation regions of western China. These dynamic seasonal distributions have wide reaching implications for waterfowl management. For instance, the areas of greatest risk for interaction between poultry and wild birds for potential zoonotic disease transmission are likely to change throughout the year, and any attempts to limit such interactions must be responsive to these changes (Prosser et al. 2013). Therefore, future projects focused on modeling disease risk should incorporate the spatio-temporal behavior and abundance of wild birds, a step our approach makes possible even in regions of limited data availability. Seasonal distributions also indicate that any successful conservation attempts should include strategies to protect the birds throughout the year, which requires attention to both wintering and breeding habitats (Cui et al. 2014; Dai et al. 2016).

While the spatio-temporal trends in avian influenza risk we observed have significant implications for how this disease should be modeled, they also have direct ecological implications. For instance, the Tufted Duck (Aythya fuligula) has been found to be especially prone to infection from highly pathogenic avian influenza, with infections often resulting in mortality (Keawcharoen et al. 2008). While the Tufted Duck is considered a species of least concern, other closely related species within the genus Aythya range from Vulnerable to Critically endangered (IUCN 2017, Table 1). Given the proclivity of some birds within the Aythya genus to suffer mortality from HPAI infections (Keawcharoen et al. 2008; Spackman et al. 2017) it is of great conservation importance to understand the risks such species face over time and space. Our findings suggest that the risk of these species to avian influenza virus is far from static, changing predictably based upon seasonal behavioral patterns. Informing our understanding of when and where vulnerable species are at greatest risk of exposure not only opens the door for more directed sampling efforts, but also enables the initiation of protective measures that could limit potential population losses. For instance, restrictions could be placed on poultry grazing and transportation through regions of seasonal importance so as to limit potential contact. Such practices could reflect the preventative measures applied in countries such as the United Kingdom (Gibbens 2017) and Thailand (Aengwanich et al. 2014) where risk to avian influenza is determined by geographic region and facility size, and biosecurity protocol are implemented accordingly.

The spatio-temporal changes in species richness observed between our wintering and breeding season models serve to highlight another key aspect of our approach: identifying areas of importance from a habitat conservation perspective. Though already established in the literature, our cumulative species richness maps confirm the importance of the Yangtze River (Cao et al. 2008, 2010; Cong et al. 2011; Williamson et al. 2013) and northeastern China (Williamson et al. 2013) in a spatial context during the wintering and breeding seasons, respectively. Additionally, our genera-specific models allow for the identification of important locations at a much finer scale, such as identifying the Changtang Plateau in Tibet as being important to birds of the genus Anser during the breeding season (Zhang et al. 2015a, b). Our models also compared well to species distribution models prepared for this region. For instance, while there were some local areas of disagreement between our models and those produced by Zeng et al. (2015), both identified the Yangtze River Basin as the region within China of the greatest probability of occurrence for the Endangered Scaly-sided Merganser. Similarly, the models produced by Dai et al. (2016) report that key Bar-headed Goose summering habitat is found at Qinghai Lake and the Tibetan Plateau, findings which concur with our results. An oft noted benefit of SDM’s is their usefulness in highlighting crucial habitats to target for conservation and intensive surveys (Moriguchi et al. 2013; Ochoa-Quintero et al. 2010). However, since all that is required to conduct our habitat suitability analysis is knowledge of a species’ habitat needs and publicly available habitat data, this method shows great promise to aid the conservation of species for which critical habitats may be less well understood, especially in these low data scenarios.

The ability of our approach to identify important habitats for waterfowl conservation is demonstrated when our model outputs are overlaid with the spatial boundaries of China’s protected areas. For instance, our outputs identify areas of suitable habitat for multiple species around Taihu Lake, near the mouth of the Yangtze River, where important wintering habitat is only sparsely protected. Our models also identify the need to focus on the protection of wintering habitats which are less protected than the breeding habitats (Cui et al. 2014). Both of these overarching trends are seen with the Ferruginous Pochard, a Near Threatened species. In China, this species breeds mostly throughout the Tibetan Plateau and the northwestern plain, with 31.9% of its suitable breeding habitat covered within expansive protected areas in this region. Conversely, the Ferruginous Pochard winters in southeastern China with only 9.7% of its suitable wintering habitat protected via a patchwork of smaller protected areas. Therefore, this species would likely most benefit from increasing protected area connectivity within its suitable wintering habitat. The information provided by our models is especially important in the context of our study, as several of the species for which we created suitability maps are recognized as species of conservation concern by the IUCN (2017), and recent literature has highlighted the important role of protected areas in species conservation (Zhang et al. 2015a, b) within regions with effective governance (Amano et al. 2017), while advocating for the addition of new protected areas within China (Cao et al. 2010; Xia et al. 2016; Xu et al. 2017). Similarly, the outputs of our model can be used to monitor changes in available habitat over time. For instance, by utilizing imagery from different years, researchers and managers can monitor changes in habitat availability for species of interest. The ability of conservationists to identify and respond to such changes is becoming increasingly important in the face of global climate change (Gillson et al. 2013; Yu et al. 2017), and models such as those created in this work provide the opportunity for responsive management even in these low data environments.

Though the comparability of our findings to other previous works provides some confidence in this approach, our formal validation provides critical additional support. The validation process demonstrated capacity to predict presence locations for individual species, however, because the validation data included presence-only records, we were not able to assess how well the models predicted the absence of a species, a common issue related to using presence-only data (Hernandez et al. 2006). Distribution models tend to be more successful in predicting presence locations than absence locations except for species with a very narrow niche (Brotons et al. 2004; Hernandez et al. 2006). We expected our models to be better at predicting presence than absence locations and recognize that this bias could translate into potential over-prediction of a species’ distribution. However, with the intent of this model being to create an input layer for disease interface modeling we felt that a more inclusive approach was appropriate as less inclusive models would run the risk of excluding potentially important wild bird-domestic poultry interfaces. This inclusivity also has benefits when utilized in a conservation context. For instance, identifying the fundamental niche may be best suited when conserving areas around presently at-risk species as this provides buffers around the current areas of use and protects valuable habitat as populations rebound and expand. Yet it is still important to recognize that our approach identifies only suitable habitat for the given species, and thus depicts much less area than species range maps. It is through this enhanced degree of selectivity that important areas can be identified and targeted by managers and researchers as demonstrated above. In any modeling exercise, the implicit limitations should be identified and the model complexity should match the objectives of the study (Merow et al. 2014), as we did here.

One aspect of our approach that warrants special attention from those who wish to apply this method in future work is the impact the selected range map can have on the final model output. The benefit of masking model outputs to species range maps is that suitable habitat only within the species known distribution is considered. This provides obvious advantages for the utility of generated products. However, range maps often represent space use trends over many years, and may not reflect current distributions. For instance, there have been recent reports of Mute Swan (Cygnus olor) wintering along the Yellow River delta (Brazil 2009), although the behavior is believed to be rare and consist of relatively few individuals. Unfortunately, observations of this behavior occurred after the publication of the range maps, chosen for their coverage of all relevant species, and as such no winter model was developed for this species. Given the rarity of this behavior, it is unlikely such an omission would have significant implications on the utility of model outputs, especially from a disease interface standpoint. Similarly, range maps make all areas of a species range appear equal in use, which can results in misleading results. For instance, White-winged Scoter (Melanitta fusca) winter predominantly along the eastern coast of China but have been known to utilize some inland habitats near Nanchong (MacKinnon and Phillipps 2000). Therefore, while inland habitat use is very limited, our model output does not differentiate the likelihood of use between these two areas of suitable habitat. This issue could likely only be resolved via more complex modeling methods which would not be supported by the data available for this region. While these limitations do not remove the benefit of masking our model outputs with range maps, researchers should carefully consider the implications of potential shortcomings, and test for errors of omission whenever datasets are sufficient.

Our need for the information gained via modeling approaches, without having access to sufficient fine scale standardized datasets, forced us to use a habitat suitability approach. However, science is always changing, and as more data is collected additional analysis options become available. Therefore, we also outlined a framework for improving our current models as new data become available through the expansion and enhancement of waterfowl monitoring during the breeding season as well as the winter via the Asian Waterbird Census and other monitoring efforts (Fig. 1b). Improvements can be made iteratively based on availability of data. For example, an intermediate level model that revises species ranges and adds points for validation could be implemented with minimal additional effort and cost, and as more comprehensive data become available newer approaches to SDM can be employed (Fig. 1b). This multi-level approach, where models are meant to be improved as new information becomes available follows principles of Bayesian logic (Jewell et al. 2009) and Adaptive Resource Management (Allan and Stankey 2009), and can be applied towards a diverse range of distribution needs. Even if this method is not ideal when compared to approaches utilized under conditions of high data availability, it does not change for science-based management. Therefore, we hope that researchers from a wide array of disciplines will recognize the opportunities this approach holds for conducting research and informing management in low data environments.