Assessing the predictive capability of randomized tree-based ensembles in streamflow modelling

Abstract. Combining randomization methods with ensemble prediction is emerging as an effective option to balance accuracy and computational efficiency in data-driven modelling. In this paper, we investigate the prediction capability of extremely randomized trees (Extra-Trees), in terms of accuracy, explanation ability and computational efficiency, in a streamflow modelling exercise. Extra-Trees are a totally randomized tree-based ensemble method that (i) alleviates the poor generalisation property and tendency to overfitting of traditional standalone decision trees (e.g. CART); (ii) is computationally efficient; and, (iii) allows to infer the relative importance of the input variables, which might help in the ex-post physical interpretation of the model. The Extra-Trees potential is analysed on two real-world case studies – Marina catchment (Singapore) and Canning River (Western Australia) – representing two different morphoclimatic contexts. The evaluation is performed against other tree-based methods (CART and M5) and parametric data-driven approaches (ANNs and multiple linear regression). Results show that Extra-Trees perform comparatively well to the best of the benchmarks (i.e. M5) in both the watersheds, while outperforming the other approaches in terms of computational requirement when adopted on large datasets. In addition, the ranking of the input variable provided can be given a physically meaningful interpretation.