Predictor complexity and feature selection affect Maxent model transferability: Evidence from global freshwater invasive species

Aim Ecological niche models (ENMs) are widely used to address urgent real-world problems such as climate change effects or invasive species; however, the generality of models when projected through space and/or time, that is transferability, remains a key challenge. Here, we explored the effects of complex predictors and feature selection on ENM transferability in a widely employed algorithm, Maxent, using five globally invasive freshwater species as case studies. Location Global. Methods We modelled the global distributions of five notorious freshwater invasive species (African sharptooth catfish Clarias gariepinus, Mozambique tilapia Oreochromis mossambicus, American bullfrog Lithobates catesbeianus, red swamp crayfish Procambarus clarkii, and Australian redclaw crayfish Cherax quadricarinatus), using three predictor datasets of varying complexities derived from two commonly used climatic data sources (WorldClim and IPCC) and three methods of model tuning that differentially incorporated feature selection. Spatially explicit transferability assessments were then conducted using a suite of evaluation metrics previously used to quantify Maxent model performance. Results We show that in the absence of detailed biological knowledge of focal species, simpler predictor datasets produce models that are more accurate than those calibrated using comprehensive bioclimatic datasets. Additionally, we find that tuning models for both optimal regularization parameters as well as feature-class combinations led to the greatest increases in transferability and geographic niche conservatism. Results indicate a tenuous link between model transferability and Akaike's information criterion corrected for small sample sizes (AICc), suggesting that the indiscriminate use of AICc as an estimate of model parsimony may lead to erratic model performance. Main conclusions Our findings demonstrate that methodological considerations can drastically affect the reliability of spatial and possibly temporal projections, which has severe implications when ENMs are used to infer species' niches, and quantify ecological or evolutionary change across impacted landscapes.

Automated summary

The study demonstrates that using simpler predictor datasets can produce more accurate models than comprehensive bioclimatic datasets when detailed biological knowledge of focal species is lacking. Additionally, tuning models for optimal regularization parameters and feature-class combinations leads to the greatest increases in transferability and geographic niche conservatism. The findings suggest a delicate balance between model transferability and AICc, cautioning against the indiscriminate use of AICc as an estimate of model parsimony for stable model performance.