Non-linear approach to grouping, dynamics and organizational informatics of benthic macroinvertebrate communities in streams by artificial neural networks

Artificial neural networks were implemented to pattern and predict benthic macroinvertbrate community in streams. Properties of self-organization, adaptability and flexibility made of artificial neural networks were networks useful for extracting information out of complex community data in various ways: grouping for classification and ordination, prediction of community dynamics, verification of environmental impacts, and revealing organizational aspects of community. Based on unsupervised learning with the Kohonen network and ART, groupings were efficiently conduced to classify and ordinate community data. The combined networks of ART and Kohonen were further utilized to group community changes. Short-time predictions of community dynamics were also possible through temporal application of artificial neural networks. The time-delayed multi-layer perceptron, and the partially and fully connected recurrent networks were able to forecast the future level of community abundance given by the previous data as input. The recurrent networks appeared to predict the temporal development of communities better. The fully connected recurrent network also effectively accommodated environmental factors, and the sensitivity analyses further revealed the impact of environmental factors on community dynamics. The organizational informatics were also patterned by artificial neural networks. Patterns of relationships among different hierarchical levels in benthic macroinvertebrate communities were effectively elucidated by the counterpropagation network. The Kohonen network and multiplayer perceptron were further utilized to characterize exergy, an integrative parameter indicating thermodynamic information in community. Temporal exergy changes were grouped by the Kohonen network and community-exergy relationships were effectively patterned by the multi-layer perceptron with the backpropagation algorithm.