Ecological modeling in two areas: species distribution modeling using commonness optimization and risk assessment for honeybees

Ecological systems are complex. Ecological models are simplified versions of ecological systems. Scientists use ecological models to better understand and manage ecological systems. I applied ecological models in two areas. The first area is the use of commonness optimization to predict species distributions in space. The second area is the use of a honeybee simulation model for risk assessment of honeybees. Humanity needs to know where species are in order to protect and manage them. Because sampling (data collection) of species is limited, knowledge of species distributions is incomplete. As a result, methods are needed to infer species distributions from sparse sampling data. Species distributions can be predicted from sparse sampling data by ecological models using commonness optimization. Commonness optimization works as follows. First, community-level estimates of species richness, the number of species per site, and dissimilarity, the differences in species between pairs of sites, must be estimated. Both estimates can then be combined to estimate the pairwise commonness, the number of species shared between pairs of sites. Pairwise commonness, in turn, can serve as an optimization target for predicting species distributions. Commonness optimization rearranges species occurrences with the goal that the commonness of the predicted community composition matches the a priori estimated pairwise commonness. Information about the prediction workflow for commonness optimization was incomplete. To address this, my co-authors and I created a how-to guide to facilitate the adoption of commonness optimization by other scientists (Chapter Two). The how-to guide covers all the steps from sparsely sampled survey data to predicted species occurrences. To be specific, we predicted bird species distributions for a tropical megacity. Because estimated richness and dissimilarity account for rare species, commonness optimization may be useful for predicting the occurrence of rare species. However, commonness optimization has not been validated for its ability to predict rare species. In addition, commonness optimization has not been compared with other ecological models. To this end, my co-authors and I validated commonness optimization for its ability to predict the occurrence of rare species (Chapter Three). We also compared the predictions of commonness optimization with those of other models. We did this by using six datasets to predict species occurrence. We found that one of the other methods, multi-label random forest, predicted species occurrence better and about 5000 times faster. Rare species were predicted poorly by commonness optimization, but better by, for example, multi-label random forest. However, the predictions made by commonness optimization had smaller richness and dissimilarity errors than the predictions made by the three other methods. Published results of commonness optimization seemed much better than I would have expected based on my own work on commonness optimization. In short, commonness optimization could correctly predict about 50% of land snail occurrences for 1350 species and about 6.7 million sites (Mokany et al., 2011). In Chapter Four, I reviewed the study design of the land snail predictions. I believe that the land snail occurrences used for validation were probably not kept separate from the predictions. This may have led to overly optimistic results in terms of correctly predicted species occurrences. In Chapter Four, I also used a commonness optimization algorithm to replicate the published predictions for synthetic species communities (Mokany et al., 2011). In short, commonness optimization was able to correctly predict approximately 100% of species occurrences for a total number of 30 species and 100 sites. My predictions successfully replicated these results. However, when I increased the total number of species from 30 to 90, the proportion of correctly predicted species occurrences dropped by half, from 100% to only 50%. These results suggest that predictions for large numbers of species and sites may not result in similarly high proportions of correctly predicted occurrences as predictions for small numbers of species and sites. Overall, my co-authors and I identified one potential strength and several weaknesses of commonness optimization. One potential strength is that commonness optimization can predict community composition with small richness and dissimilarity errors (Chapter Three). This may be useful for predictions where there is a strong emphasis on plausible dissimilarity, but little emphasis on correct species identities. An example is the prediction of initial community composition for simulation studies where all species are assumed to be similar. However, we have also found weaknesses in commonness optimization. First, other models predict species occurrences predict species occurrences better and much faster (Chapter Three). Second, geographically rare species are poorly predicted by commonness optimization (Chapter Three). Because such rare species have little effect on the overall commonness, commonness optimization is unlikely to correctly predict the occurrence of rare species. Third, commonness optimization is slow, as predictions for many sites and species can take weeks. Long computation times may make the future development and improvement of commonness optimization difficult. Risk assessment for honeybees is another useful application of ecological models. Honeybees are important pollinators, and contribute to the global food production. However, honeybees are exposed to multiple stressors. A first example is agrochemicals such as the insecticide imidacloprid. A second example is landscape simplification. Landscape simplification leads to temporal gaps in the availability of nectar and pollen around the honeybee hives. A third example is exposure of honeybees to varroa mites, which transmit deformed wing virus. BEEHAVE (Becher et al., 2014) is a model of a honeybee colony. In short, the BEEHAVE model consists of a simulated honeybee colony with a queen, foragers, workers, drones, and larvae. It is designed to capture the colony dynamics inside the hive and foraging behavior outside the hive. However, the effects of insecticides on honeybees were missing from BEEHAVE. My co-authors and I extended BEEHAVE to include chronic (long-term) and acute (short-term) effects of imidacloprid. We exposed the honeybee colony to three stressors. First, imidacloprid originating from flowering oilseed rape. Second, periods when pollen and nectar were not available. Third, varroa mites carrying the deformed wing virus. We found few acute effects of imidacloprid because the thresholds for such effects were rarely, if ever, reached. However, we found that colony performance was affected by chronic effects. Colony performance was more affected in the second and third years of exposure than in the first year. Our study was the first extension of the BEEHAVE model (Becher et al., 2014) to include the effects of imidacloprid. First, it demonstrates the importance of long-term studies of pesticide effects. Second, it represents a successful application of ecological modeling in the field of risk assessment and food safety.

2024-07-16