Effect of sample size on habitat suitability estimation using random forests: a case of bluegill, Lepomis macrochirus

Abstract

Species distribution models (SDMs) have been used to understand the habitat suitability of key species. Habitat suitability plots, one outcome from SDMs, are valuable for understanding the habitat suitability and behavior of organisms. The sample size is often constrained by budget and time, and could largely influence the reliability of habitat suitability plots. To understand the effect of sample size on habitat suitability plots, the present study utilized random forests (RF) combined with partial dependence function. And the bluegill (Lepomis macrochirus), a main exotic fish species in the Japan rivers, was selected as target species in this study. Total of 1010 samples of bluegill observations along with four environmental variables were surveyed by the National Censuses on River Environments. The area under curves was calculated after generating RF models, to assess the predictive model performance, and this process was repeated 1000 times. To draw habitat suitability plots, we applied partial dependence function to the formulated RF models, and 15 different sample sizes were set to examine the effect on habitat suitability plots. We concluded that habitat suitability plots are affected by sample size and prediction performance. Notably, habitat suitability plots drawn from the sample size of 50 greatly varied among the 1000-time iterations, and they are all different from the observations. Furthermore, to deal with the case of limited samples, we proposed a novel approach “averaged habitat suitability plot” for delineating habitat suitability plots. The proposed approach enables us to assess the habitat suitability even with a small sample size.