Invasive bighead carp (Hypophthalmichthys nobilis) and silver carp (Hypophthalmichthys molitrix) established in the lower Mississippi River and continue to spread into the Upper Mississippi Basin and threaten to invade the Laurentian Great Lakes. We sought to estimate the potential distribution of these species in North America using seven machine learning (ML) classifiers, which include instance-based, statistical, and second generation classical ML approaches. First, we developed models trained using native range information. For accuracy assessments, we determined the number of non-native occurrence information classified correctly for each model. We applied the classifiers to North American climate information to generate an estimate of the potential geographic extent for each species. Next, we incorporated subsets of non-native occurrence information into the training set and estimated the change in accuracy and extent for 10, 15, 20, 25, and 50 locations from the set of established populations. Effects of geographic extent with additional data and for similarity between the geographic patterns of classification were tested with statistical models.
Results/Conclusions
Models trained with bighead carp native range information correctly classified subsets of established populations regardless of method. Models for silver carp were more accurate, especially Self-Organizing Maps and two instance-based approaches. Occurrence information from non-native, established populations increased accuracy and reduced variability for all methods and both species. Most silver carp models correctly classified all established populations with 10 more points; bighead carp models showed continual improvement with the additional data (up to 50 points). We found that geographic extent increased with additional data for five of seven methods and one method for bighead and silver carp, respectively. However, the geographic patterns of classification were similar and slope coefficients for models generally were small given the overall extent of projection. When we compared geographic patterns between methods, estimates of distribution were not similar even where accuracy values were similar. Our results show that if a limited number of native range occurrence information is available, a limited amount of non-native range can improve performance. Furthermore, by implementing several machine learning techniques to the problem of invasive species distributions, we show that model choice results in a sensitivity in the expectation of geographic extent.