Informatics software for the ecologist's toolbox: A basic example

Abstract

Machine learning techniques for ecological applications or “eco-informatics” are becoming increasingly useful and accessible as software for these techniques becomes more readily available. Complex ecological data sets with a multitude of variables are also increasingly available. Ecologists, who do not necessarily have extensive backgrounds in machine-learning techniques, are facing decisions on new methods of data analysis. We evaluated the predictive ability of three commercially available (i.e. user-friendly) software packages for artificial neural networks (ANNs), evolutionary algorithms (EAs), and classification/regression trees (CART). To demonstrate their usage, we analyzed fish and habitat data from the mid-Atlantic region of the US, which was collected by the U.S. Environmental Protection Agency (EPA). These data, including over 200 environmental descriptors summarizing watershed, stream, and water chemistry, and physical habitat characteristics in addition to fish community metrics (i.e. richness, Index of Biotic Integrity (IBI) scores, % exotics), were collected as part of the EPA's Environmental Monitoring and Protection program. We predicted fish IBI scores as a function of these local and regional scale habitat variables. Predictive ability is evaluated with independent validation data. These approaches could prove especially useful for conservation or management applications where ecologists seek to utilize the most comprehensive data to make predictions at various scales. By employing “user-friendly” software we hope to show that ecologists, without extensive knowledge of computational science, can benefit from these techniques by extracting more information about complex ecosystems. We found that all models predicted better than chance ( p < 0.05). Relative strengths and weaknesses of these three approaches are compared and recommendations for their use in ecological applications are presented.