Abstract
Machine learning (ML) techniques have become important to support decision
making in management and conservation of freshwater aquatic
ecosystems. Given the large number of ML techniques and to improve the
understanding of ML utility in ecology, it is necessary to perform comparative
studies of these techniques as a preparatory analysis for future model
applications. The objectives of this study were (i) to compare the reliability
and ecological relevance of two predictive models for fish richness, based
on the techniques of artificial neural networks (ANN) and random forests
(RF) and (ii) to evaluate the conformity in terms of selected important variables
between the two modelling approaches. The effectiveness of the
models were evaluated using three performance metrics: the determination
coefficient (R2), the mean squared error (MSE) and the adjusted determination
coefficient (R2adj) and both models were developed using a k-fold
crossvalidation procedure. According to the results, both techniques had
similar validation performance (R2 = 68% for RF and R2 = 66% for ANN).
Although the two methods selected different subsets of input variables,
both models demonstrated high ecological relevance for the conservation
of native fish in the Mediterranean region. Moreover, this work shows how
the use of different modelling methods can assist the critical analysis of
predictions at a catchment scale.