Algae-based Biomonitoring: Predicting Diatom Reference Communities in Unpolluted Streams using Classification Trees, Random Forests, and Artificial Neural Networks

Abstract

Abstract The Eastern Canadian Diatom Index (IDEC) was developed to evaluate the ecological integrity of streams along a pollution gradient, as a function of the dissimilarity between current diatom communities and suitable reference communities. Distinguishing natural variations in community structure from those induced by human activities is essential for proper assessment of dissimilarity. To account for the effect of the natural variation in pH on this assessment, two IDEC subindices were used: one for sites with diatom reference communities typical of naturally alkaline water pH, and another for sites with communities typical of naturally circumneutral water pH. This study used three statistical models, namely classification trees (CT), random forests (RF), and artificial neural networks (ANN) to: (i) identify the environmental variables discriminating between alkaline and neutral reference communities (“biotypes”), and (ii) compare their predictive capacities. Models identified clay rocks, gneiss/paragneiss rocks, siliceous rocks, and carbonated rocks as the main geological features discriminating reference biotypes. For the reference streams, clay, siliceous, and carbonated rocks were associated with high water pH while gneiss/paragneiss rocks were associated with low water pH. Both ANN and RF models behaved similarly across all performance criteria and yielded general models useful for identifying the appropriate IDEC sub-index.