Wetland bioassessment using Naïve Bayesian indicator species analysis

Background/Question/Methods

The use of ecological indicators for bioassessment has been rapidly increasing in recent years. Ecological indicators aid assessment of ecological condition to inform regulatory, stewardship, sustainability, or biodiversity decisions. The presence or absence of species themselves has long been used as indicators of habitat type or condition. Because disturbance can both eliminate species that would often be present in an undisturbed site, and allow species to invade a site that they would otherwise be unable to, information on species presence/absence can be used to create assessment tools to quantify the impact of human disturbance. However, since there are multiple factors other than disturbance that can contribute to the presence/absence of a species at a site, accurate use of indicator species requires modeling the joint effects of both natural environmental gradients and human disturbance. Furthermore, because of stochasticity inherent in the presence\absence of species, it is desirable to combine information from multiple indicator species to increase the accuracy of the predictions. Here, we discuss a bioassessment model that addresses these needs.  We develop a Naïve Bayesian multi-indicator species model and apply it to assess the condition of riparian wetland in Rocky Mountain National Park.

Results/Conclusions

We applied this method to wetland data from the National Park Service’s Inventory and Monitoring Program in Rocky Mountain National Park. The data include information on 89 plant species across 38 riparian wetland sites and a suite of covariates. Using leave-one-out cross-validation, we modeled the probability of the presence each species as a function of human disturbance and elevation. The resulting parameter estimates were combined with the held-out data on the presence/absence of the species, environmental factors and an uninformative prior to quantify a posterior predictive distribution of the level of human disturbance at the test site. We found that the modes of the predicted posterior distributions were highly correlated with the actual value of human disturbance (r=0.78). These results suggest that Naïve Bayesian multi-indicator species models can be used to make robust assessments of the degree of human impact biological systems and could be a useful tool for bioassessment.