The structure of Collembolan assemblages in two Pyrenean springs (Ruau and La Maure) was analyzed using Self-Organizing Map (SOM). The dataset included 84 samples cores were taken from each spring. Samples were collected every two months from December 1993 to December 1994. 59 species of Collembola were collected, of which 53 were present at Ruau and 43 at La Maure.

From the species dataset, the community assemblage groups were defined by the SOM, indicating an unexpected heterogeneity. The trained mapping by the SOM showed patterns of communities from different substrates and different temperature conditions, e.g. substrate, water content, temperature...

Then, we evaluated the predictive ability of predictive models of artificial neural networks (backpropagation algorithms), evolutionary algorithms, and trees models (CART & Random forest) to predict community assemblages and biodiversity. The contribution of variables in the models was also established.

The present results suggest that the training by the present artificial neural network can be an effective tool for organizing community data in a large-scale ecological survey.