Invaders Main | Noxious Weeds List | Links Database | Queries by: Name, List, Area, and Map

Project Map

Data Sets
     Training Set
     Prediction Set
     Initial Experiments
     Neural Network

        Designation of a plant as a noxious weed under state law may be the most important and complex decision made by vegetation managers. The process of regulatory listing is highly dependent on the advice of expert panels. However, experts advice is constrained by the large number of introduced exotics, incomplete data, and the limitations of individual experience and knowledge of recent and new invaders. Noxious listing also tends to be reactive rather than proactive. Many weeds are recognized as noxious only after they have become widespread in their newly invaded region. Late listing precludes effective eradication and containment goals. Widely established weeds have perpetual impact and suppression costs. An objective numerical approach to classifying new invaders as noxious would be an asset to the regulatory process.

        Weed ecologists consider certain plant traits and bio-geographic factors to be associated with invasive behavior. A data set of 21 invasive attributes was constructed for all 554 exotics believed to have established outside artificially maintained settings in Idaho and Montana. The exotics were split into the 434 species that had arrived prior to 1951, and the 120 reported since 1951. Exotics arriving before 1951 were used to train, cross verify, and validate several types of artificial intelligence algorithms that classified the plants as noxious or not.

        The best performing model was a multilayer perceptron neural network. Five input attributes contributed to the optimal model:

  • Native European/Asian latitudinal range
  • Absolute latitudinal range where native and exotic in Europe/Asia
  • The number of counties reporting infestations
  • How many states or provinces outside of Idaho and Montana have declared the plant noxious
  • The life form of the plant (forb, shrub, tree, grass-like, fern)

        The selected neural network model had good overall performance. The classification rate for the model training dataset was 95% correct. Cross verification error was only 23% indicating that the model was not over-trained to the data and could generalize. The exotic species used to test or validate the model were 83% correctly classified.

        The 120 exotic species reported as arriving since 1951 were then analyzed with the selected neural network. Twenty-nine post-1950 arrivals were classified as noxious, including 11 of the 16 post-1950 introductions that have already been legally designated as noxious by one or more Pacific Northwest states.

        Although the artificial neural network approach preformed quite well according to criteria used to evaluate machine learning algorithms, the current version of the risk assessment model can only make a limited contribution the regulatory process beyond that which is being provided by expert panels. The training dataset and neural network model may be improved by further effort. Exploratory improvements include:

  • Removing anomalous cases from the training dataset
  • Increase the number of training cases with complete attribute data
  • Obtain more biological attributes associated with weedy behavior (seed size, plant height, etc.)

        Weeds new to the region can be quickly analyzed as single cases. Established weeds can also be easily reanalyzed as attributes, such as number of counties reporting infestations, change. The attribute data can be viewed from the within this report on the INVADERS web site.

Introduction >>