A modeling workflow that balances automation and human intervention to inform invasive plant management decisions at multiple spatial scales

Abstract

Predictions of habitat suitability for invasive plant species can guide risk assessments at regional and national scales and inform early detection and rapid-response strategies at local scales. We present a general approach to invasive species modeling and mapping that meets objectives at multiple scales. Our methodology is designed to balance trade-offs between developing highly customized models for few species versus fitting non-specific and generic models for numerous species. We developed a national library of environmental variables known to physiologically limit plant distributions and relied on human input based on natural history knowledge to further narrow the variable set for each species before developing habitat suitability models. To ensure efficiency, we used largely automated modeling approaches and human input only at key junctures. We explore and present uncertainty by using two alternative sources of background samples, including five statistical algorithms, and constructing model ensembles. We demonstrate the use and efficiency of the Software for Assisted Habitat Modeling [SAHM 2.1.2], a package in VisTrails, which performs the majority of the modeling analyses. Our workflow includes solicitation of expert feedback on model outputs such as spatial prediction results and variable response curves, and iterative improvement based on new data availability and directed field validation of initial model results. We highlight the utility of the models for decision-making at regional and local scales with case studies of two plant species that invade natural areas: fountain grass (Pennisetum setaceum) and goutweed (Aegopodium podagraria). By balancing model automation with human intervention, we can efficiently provide land managers with mapped predicted distributions for multiple invasive species to inform decisions across spatial scales.

Fig 1. Workflow of the modeling framework showing data sources, model input data, automated and human processes, model output products and the paths for model iterations.

Fig 2

A-F: Potential habitat suitability model ensemble (maximum value of 10) using the one percentile threshold for fountain grass (Pennisetum setaceum) (A,C,E) and goutweed (Aegopodium podagraria) (B,D,F) at each extent: national extent (A, B), regional extent defined by the Exotic Plant Management Team Regions including C) Lake Mead and D) Great Lakes), and the local extent defined by E) Joshua Tree National Park and F) Pictured Rocks National Lakeshore.

Fig 3. Model ensemble values associated with the independent observation data for fountain grass from CalFlora for four different threshold metrics including minimum predicted presence, one percentile, ten percentile and the maximum of sensitivity plus specificity.

Fig 4. Fountain grass (Pennisetum setaceum) models of four different thresholds including minimum predicted presence (MPP), one percentile, ten percentile and maximum of sensitivity plus specificity (MSS).

Model predictions are shown at three scales; A) national, B) regional (Lake Mead Exotic Plant Management Team region in blue) and C) local (Joshua Tree National Park in green).