Aquatic macrophytes are an undervalued but critically important component of Minnesota’s lakes. Recent development of a multimetric macrophyte index (MMI) suggests that aquatic plants are useful indicators of lake condition. A quantitative evaluation of the MMI to determine whether the index can explicitly link lake condition with activities that negatively impact lake resources has not been conducted. Additionally, the statistical properties of multimetric indices may confound the interpretation of index response to changes in aquatic habitat condition. A greater understanding of index performance is necessary before it can be used to accurately and precisely evaluate lake health. Analytical approaches using ecological informatics could provide a powerful technique to evaluate the MMI. Supervised neural network models may be particularly useful given the ability to characterize multivariate response and relaxation of assumptions necessary for more conventional modeling techniques. The objective of this study was to develop supervised neural networks to quantify MMI response to environmental variables related to lake condition. These models were expected to illustrate key predictors of MMI performance with implications for lake management. Additionally, we hypothesized that neural networks could provide a general technique to evaluate multimetric index performance in other systems or regions.
Results/Conclusions
Neural networks made precise predictions of overall MMI scores using an independent dataset (R2 fit between observed and predicted scores was 0.61). Predictive performance of the neural network models varied for individual metrics (R2 from 0.12 – 0.68). Bootstrap analyses to evaluate the effects of different training data on model performance indicated that predictions were highly sensitive to the training data. However, consistent relationships were observed between specific metrics and explanatory variables, such as maximum depth of plant growth with trophic state. Metrics that were consistently related to specific explanatory variables were further evaluated to quantify relationships in the context of other explanatory variables. Overall, the neural networks did not identify the specific relationships between the explanatory variables and metrics of the MMI, although some consistent relationships were identified. More conventional modeling techniques had similar performance in predicting MMI and metric response, suggesting the MMI inherently confounds relationships among variables, rather than an inability of neural networks to characterize variable importance. These results suggest that the statistical properties of multimetric indices should be carefully evaluated during index development, with specific attention given to the sensitivity and unique diagnostic capabilities of individual metrics.