Modelling fish co-occurrence patterns in a small spring-fed river using a machine learning approach

Abstract

Stopping biodiversity loss due to human activities is of great concern for ensuring ecosystem services for sustainable development. While species-habitat relationships are an important aspect of species conservation, considering species interactions remains challenging in habitat suitability modelling due to the complex nature of the species’ ecology and uncertainties present in ecological data. To cope with this challenge, a Random Forests (RF) multi-class classifier is built using a data set comprised of monthly fish habitat surveys conducted over a period of two years for analyzing the 16 co-occurrence patterns among four fish species, including endangered species and translocated domestic alien species, in a small spring-fed river in Japan. This multi-class habitat suitability model allows for assessing the instream habitat conditions under which multiple species co-occur. Variable importance and partial dependence plots are generated to extract ecological information for a deeper understanding of the species co-occurrence patterns. Results show that the model performance is greatly influenced by data prevalence as reported for single-species habitat suitability models. Ecological information extracted from the RF model illustrates instream habitat conditions that are important for each of the co-occurrence patterns. Such information can be used for prioritizing conservation sites and better designing habitat conditions in order for multiple native species to co-occur or for preventing non-native species to expand their distribution ranges.

Keywords

• Freshwater fish
• Habitat suitability
• Habitat conservation and management
• Partial dependence plots
• Variable importance
• Random Forests