Incorporating ecological knowledge into ecoinformatics: An example of modeling hierarchically structured aquatic communities with neural networks

The field of ecoinformatics is concerned with gaining a greater understanding of complex ecological systems. Many ecoinformatic tools, including artificial neural networks (ANNs), can shed important insights into the complexities of ecological data through pattern recognition and prediction; however, we argue that ecological knowledge has been used in a very limited fashion to shape the manner in which these approaches are applied. The present study provides a simple example of using ecological theory to better direct the use of neural networks to address a fundamental question in aquatic ecology—how are local stream macroinvertebrate communities structured by a hierarchy of environmental factors operating at multiple spatial scales? Using data for 195 sites in the western United States, we developed single-scale, multi-scale and hierarchical multi-scale neural networks relating EPT (Orders: Ephermeroptera, Plecoptera, Trichoptera) richness to environmental variables quantified at 3 spatial scales: entire watershed, valley bottom (100s–1000s m), and local stream reach (10s–100s m). Results showed that models based on multiple spatial scales greatly outperformed single-scale analyses (R = 0.74 vs. R¯ = 0.51) and that a hierarchical ANN, which accounts for the fact that valley- and watershed-scale drivers influence local characteristics of the stream reach, provided greater insight into how environmental factors interact across nested spatial scales than did the non-hierarchical multi-scale model. Our analysis suggests that watershed drivers play a greater role in structuring local macroinvertebrate assemblages via their direct effects on local-scale habitats, whereas they play a much smaller indirect role through their influence on valley-scale characteristics. For the hierarchical model, the strongest predictors of EPT richness included descriptors of climate, land-use and hydrology at the watershed scale, land-use at the valley scale, and substrate characteristics and riparian cover at the reach scale. In summary, our results highlight the importance of incorporating environmental hierarchies to better understand and predict local patterns of macroinvertebrate assemblage structure in stream ecosystems. More generally, our case study serves to emphasize how incorporating prior ecological knowledge into ANN model structure can strengthen the relevance of ecoinformatic techniques for the broader scientific community.