Predicting weed distribution at the landscape scale: using naturalized Brassica as a model system.
Quantifying and predicting the distribution of naturalized plant species is a major concern for weed risk assessment and understanding the potential impacts of gene-flow between crop species and feral or naturalized populations. We developed a rapid field assessment to quantify the distribution of naturalized Brassica populations at the landscape scale, using the Canterbury Plains region of New Zealand as a model system. Internationally, brassicas are one of the most widely cultivated crops, and have well-documented problems of crop escapes, gene-flow and hybridization with wild relatives. Brassicas are cultivated intensively in the study system for both seed production and forage, are widely naturalized and have raised concerns for gene-flow from feral populations to cultivated varieties. Generalized linear models (GLM) were constructed in order to predict the presence-absence of brassicas using survey data of naturalized populations located in 50 3×3 km plots. This model was validated subsequently both by ground-truthing predicted presence-absence in a second field survey, and by comparing cumulative probability distributions from a presence-only botanical survey. The best model produced from the field survey predicted naturalized Brassica presence-absence robustly (r2=0.52) and was well validated against results from the presence-only survey carried out in the same year. Results from the second field survey during a different growing season did not match predictions from the original model, suggesting that strong between-year differences influence large-scale Brassica distribution. The differences between years were driven by an overall increase in the prevalence of brassicas over the course of the study that was not associated with any predictive variable or climate. Synthesis and applications. These results have important implications both for developing models of species distribution and for predicting gene-flow risk at the landscape scale. Despite constructing a model that was subsequently validated from presence-only data in one year, it did not predict Brassica distribution in a subsequent year; this suggests that between-year or stochastic factors operate strongly and are a major consideration for risk assessment. Models used to predict species distributions, in general, need to account for these between-year effects either by incorporating mechanistic processes or long-term data.