Consistent accuracy of the Australian weed risk assessment system across varied geographies

The ecological and economic advantages of preventing introduction of species likely to become invasive have increased interest in implementing effective screening tools. We compared the accuracy of the Australian Weed Risk Assessment (WRA) system with that across the six geographies in which it has been tested (New Zealand, Hawaii, Hawaii and Pacific Islands, Czech Republic, Bonin Islands and Florida). Inclusion in four of the tests of a secondary screening tool, developed to reduce the number of species requiring further evaluation, decreased the number of species with that outcome by over 60% on average. Averaging across all tests demonstrated that the WRA system accurately identified major invaders 90%, and non-invaders 70%, of the time. Examined differently, a species of unknown invasive potential is on average likely to be correctly accepted or rejected over 80% of the time for all of these geographies when minor invaders are categorized as invasive. Whereas increasing consistency in definitions and implementation would facilitate understanding of the general application of the WRA system, we believe that this tool functions similarly across islands and continents in tropical and temperate climates and has been sufficiently tested to be adopted as an initial screen for plant species proposed for introduction to a new geography.     

Predicting invasions by woody species in a temperate zone: a test of three risk assessment schemes in the Czech Republic (Central Europe)

To assess the validity of previously developed risk assessment schemes in the conditions of Central Europe, we tested (1) Australian weed risk assessment scheme (WRA; Pheloung et al . 1999); (2) WRA with additional analysis by Daehler et al . (2004); and (3) decision tree scheme of Reichard and Hamilton (1997) developed in North America, on a data set of 180 alien woody species commonly planted in the Czech Republic. This list included 17 invasive species, 9 naturalized but non-invasive, 31 casual aliens, and 123 species not reported to escape from cultivation. The WRA model with additional analysis provided best results, rejecting 100% of invasive species, accepting 83.8% of non-invasive, and recommending further 13.0% for additional analysis. Overall accuracy of the WRA model with additional analysis was 85.5%, higher than that of the basic WRA scheme (67.9%) and the Reichard–Hamilton model (61.6%). Only the Reichard–Hamilton scheme accepted some invaders. The probability that an accepted species will become an invader was zero for both WRA models and 3.2% for the Reichard–Hamilton model. The probability that a rejected species would have been an invader was 77.3% for both WRA models and 24.0% for the Reichard–Hamilton model. It is concluded that the WRA model, especially with additional analysis, appears to be a promising template for building a widely applicable system for screening out invasive plant introductions.     

Performances of machine learning algorithms for mapping fractional cover of an invasive plant species in a dryland ecosystem

In recent years, an increasing number of distribution maps of invasive alien plant species (IAPS) have been published using different machine learning algorithms (MLAs). However, for designing spatially explicit management strategies, distribution maps should include information on the local cover/abundance of the IAPS. This study compares the performances of five MLAs: gradient boosting machine in two different implementations, random forest, support vector machine and deep learning neural network, one ensemble model and a generalized linear model; thereby identifying the best‐performing ones in mapping the fractional cover/abundance and distribution of IPAS, in this case called Prosopis juliflora (SW. DC.). Field level Prosopis cover and spatial datasets of seventeen biophysical and anthropogenic variables were collected, processed, and used to train and validate the algorithms so as to generate fractional cover maps of Prosopis in the dryland ecosystem of the Afar Region, Ethiopia. Out of the seven tested algorithms, random forest performed the best with an accuracy of 92% and sensitivity and specificity >0.89. The next best‐performing algorithms were the ensemble model and gradient boosting machine with an accuracy of 89% and 88%, respectively. The other tested algorithms achieved comparably low performances. The strong explanatory variables for Prosopis distributions in all models were NDVI, elevation, distance to villages and distance to rivers; rainfall, temperature, near‐infrared and red reflectance, whereas topographic variables, except for elevation, did not contribute much to the current distribution of Prosopis. According to the random forest model, a total of 1.173 million ha (12.33% of the study region) was found to be invaded by Prosopis to varying degrees of cover. Our findings demonstrate that MLAs can be successfully used to develop fractional cover maps of plant species, particularly IAPS so as to design targeted and spatially explicit management strategies.