MODELLING HABITAT SUITABILITY OF LEFUA ECHIGONIA USING RANDOM FORESTS

Abstract

 This study assessed microhabitat conditions of Lefua echigonia using field observed ecohydraulic data and a data-driven habitat model. A series of monthly field surveys were conducted in a spring-fed small urban stream in Tokyo, Japan. Random forests (RF), a predictive machine learning method, was applied as a binary classification tool to analyze the relationship between physical habitat conditions and the presence/absence of L. echigonia. From the RF-based habitat model developed, variable importance and response curves were extracted for a deeper understanding of the ecology of target species.As a result, variable importance suggested the importance of water velocity for presence/absence modelling. Response cures illustrated the important instream habitat conditions such as shallow water with low water velocity, and high water velocity with a large percent coverage of large-sized gravels. The habitat information obtained in this study can be used to identify potential habitats for L. echigonia.