Drake Lab - Risk classification

Identifying traits of invasive species is both an intellectual challenge and a practical problem. This study aims to understand how species' individual biological traits affect their propensity for becoming invasive. Of course, the notion of invasiveness is vague, lumping together several properties of invasive species that need to be considered individually, e.g., a propensity for establishing viable populations or a propensity for being a nuisance once established. Our research aims to identify what these properties are and identify their biological and ecological underpinnings. Such correlations generally have a basis in ecological theory. For instance, the brain size-environmental change hypothesis of Sol et al. (2005) suggests that establishment success should increase with species' relative brain sizes. Alternatively, classic island biogeography theory holds that establishment success should correlate with potential reproductive output. Our analysis of establishment success in fish introductions failed to find evidence for either of these postulated patterns (Drake 2007). Instead, we identified a new predictor or establishment success-the degree of parental investment in the development of offspring (Drake 2007). In a related project with Reuben Keller we sought to identify traits of mussels that predicted whether they would be a nuisance if established (Keller et al. 2007). The result of this study boils down to one simple conclusion: environmental impact of mussels is directly related to maximum individual fecundity. Ongoing work will extend these results with new models and for different species groups. Particularly, we are aiming to develop more accurate models using computational methods that are more sensitive to nonlinear relationships. In the future we hope to apply these methods to weed classification.

We are in the midst of a project (funded by the Economic Research Service of the USDA) to develop cost-sensitive decision support tools (classification algorithms and visual decision trees) and parameterize them with empirical data to aid risk analysis for newly reported imported plant species and species proposed for future introduction. In contrast to previous studies, we are incorporating expected costs in algorithm identification to minimize expected damages rather than total errors. To support these objectives, we have (1) developed databases of species and genera introduced into the continental U.S. and Hawaii; (2) are developing theoretical and empirical models for cost/benefit distributions of weeds; and (3) using nonparametric distribution estimation techniques and unsupervised learning algorithms to detect and discriminate classes of weeds with respect to mode and magnitude of impact and biological features. To accomplish our overall goal, we will apply machine learning algorithms (neural nets, kernel-based learning algorithms, distribution estimation, nearest neighbor generalization, dissimilarity metrics, etc.) implementing techniques for variable selection and model combination to reduce complexity and dependence on data that are difficult to obtain while increasing accuracy.

The species database includes balanced pairs of invasive and non-weedy congeners (237 pairs) and traits associated with each species. (Following records available on the Plants National Database (PND), species is “invasive” if it is listed or legislated against as a “noxious weed” by the federal government or any state. Non-weeds are species that are neither listed as weeds by state of federal government nor have been reported as weedy by plant or agricultural specialists belonging to universities or government agencies.) Trait data has been compiled from PND and numerous other freely accessible sources (including the Flora of North America and regional floras) to provide data on life history, basic morphology and physiology, and habitat and geographic origins. Because in many cases, invasive species are the only member of their genus present within the U.S. or all members of the genus which have been introduced are weedy, we have also compiled a database of 1528 genera which contain successfully established introduced species. For each genus, the database of genera tallies (based on PND) the number of introduced species and the number of weedy species, and a set of genus-aggregated traits. Both databases are to be used for this study and to be archived in publicly accessible data repositories for benchmarking future developments. As a means of improving the value of the databases for current and subsequent analyses, we are currently exploring imputation techniques which will allow us to utilize variables for which a fraction of the data is missing.