DISTRIBUTION MODELS OF STREAM INVERTEBRATES CONSIDERING RIFFLE-POOL HYDRAULIC VARIATIONS USING A MACHINE LEARNING TECHNIQUE

Abstract

 Habitat models that predict the distribution of stream animals along environmental gradients at the catchment scale are useful in river environment management. However, past studies have not examined such prediction model along a broader gradient of riffle-pool hydraulic variables such as water depth. In this study, we developed a model for predicting distribution of stream invertebrates along gradients of rifle-pool hydraulic variables using an unsamble machine learning technique, random forest, in the Omaru River catchment, located in southern Japan. As a result, many biological and environmental variables (e.g., slope) significantly increased in riffle reaches compared with those in pool reaches. In the prediction models of 156 taxa that appeared at three or more locations, elevation and catchment area were the most important predictors, followed by hydraulic variables such as current velocity and water depth. As a result of performing 4-fold cross-validation by adding hydraulic variables to predictor variables, a predictive performance metric of presence / absence (i.e., AUC; Area Under Curve) significantly increased for the four orders (ephemeroptera, plecoptera, trichoptera and diptera) and five life types (burrowers, case-bearers, crawlers, net-spinners and swimmer) of stream invertebrates.