Input variable selection for ecological modelling in inter-basin water transfer management

Abstract

Inter-basin water transfers, usually driven by purely economic purposes can have complex physical and biological implications both in the upstream and downstream reservoirs. These transfers are well studied from a water quantity management point of view, but increasing pressure on water resources is encouraging decision-makers to move towards a more holistic management strategy, accounting for both quantity and quality issues. The integration of high fidelity water quality modelling within an optimization based decision-making framework calls for new mathematical tools that balance computational efficiency and accuracy. In particular, input variable selection techniques represent the first step towards low-order emulation of physically based model by selecting, among all the possible candidate inputs, the ones that better reproduce the behavior of a specified variable. In this study, a novel data-driven approach to input variable selection is applied to a 9-year dataset of phytoplankton data measured in the receiving end of two interconnected reservoirs. Preliminary analyses show that diatom growth in the receiving reservoir is influenced not only by internal processes (e.g. net growth as function of nutrients concentrations), but also by the input of algal cells from the upstream reservoir (e.g. seeding effect of particular algal species) via water transfers. To evaluate the relative importance of all the candidate input variables in explaining the output behavior, we resorted to the Iterative Input Selection (IIS) algorithm. The algorithm, which is based on Extremely Randomized Trees (Extra-Trees) as the underlying model class, incrementally builds the set of variables by adding the most significant ones according to a ranking procedure, based on a statistical measure of significance that accounts for non-linear dependencies. The IIS algorithm stops selecting new variables when the desired accuracy is achieved. The capability of IIS in selecting the most relevant input variables is shown to be superior to common methods based on cross-correlation analyses. This study represents the first application of the IIS algorithm to the interpretation of ecological data.