Predicting hotspots for invasive species introduction in Europe

Kevin Schneider,David Makowski,Wopke Van Der Werf

doi:10.1088/1748-9326/ac2f19

Abstract

Plant pest invasions cost billions of Euros each year in Europe. Prediction of likely places of pest introduction could greatly help focus efforts on prevention and control and thus reduce societal costs of pest invasions. Here, we test whether generic data-driven risk maps of pest introduction, valid for multiple species and produced by machine learning methods, could supplement the costly species-specific risk analyses currently conducted by governmental agencies. An elastic-net algorithm was trained on a dataset covering 243 invasive species to map risk of new introductions in Europe as a function of climate, soils, water, and anthropogenic factors. Results revealed that the BeNeLux states, Northern Italy, the Northern Balkans, and the United Kingdom, and areas around container ports such as Antwerp, London, Rijeka, and Saint Petersburg were at higher risk of introductions. Our analysis shows that machine learning can produce hotspot maps for pest introductions with a high predictive accuracy, but that systematically collected data on species’ presences and absences are required to further validate and improve these maps.

Highlights

Biological invasions describe inadvertent introductions of organisms into new territories
Global georeferenced data on a wide range of potential predictors related to climate, soils, water, and anthropogenic factors were collected, and an elastic-net machine learning algorithm was trained on around 341 000 observations across the globe to predict new introduction of invasive species as a function of the predictors
Our analysis shows that machine learning can produce hotspot maps for plant pest introduction with a high predictive accuracy, but that systematically collected data on species’ presences and absences are required to further validate and improve these maps

Summary

Introduction

Biological invasions describe inadvertent introductions of organisms into new territories. A generic approach that could help to identify areas that are generally more at risk for pest introduction, without having to first develop a range of species-specific models, would greatly improve evidence-based prevention and management. GBIF data usually come as presence-only and it was necessary to generate background data representing pseudo-absences to train and test our models, as commonly done in SDMs [4]. While this is common practice in the SDM literature, there is no consensus regarding the best approach [70]. Higher values for drought severity, seasonal water variability, elevation, and the land cover classification for moss and lichen as well as cultivated and managed cropland decreased risk, while the biome classification for temperate sclerophyll woodland and shrubland, and higher values for flood occurrences, biodiversity intactness, organic carbon content in the soil, average photosynthetically active radiation in September, and soil water content were associated with increased risk

Objectives

Methods

Results

Conclusion