Predicting vector distribution in Europe: at what sample size are species distribution models reliable?

Area of Science:

• Ecological modeling
• Epidemiology
• Machine learning applications

Background:

• Species distribution models (SDMs) predict disease vectors using environmental data.
• Climate change and rising vector-borne diseases necessitate improved surveillance in Europe.
• Current SDM practices lack standardization, particularly regarding optimal sample size.

Purpose of the Study:

• To determine the optimum sample size for Random Forest models.
• To evaluate the impact of different sample ratios on model reliability.
• To inform standardized practices for vector distribution modeling.

Main Methods:

• A simulated vector with a known distribution was used across 10 European test sites.
• 9,000 Random Forest models were trained with varying sample sizes (10-5,000) and ratios (50:50, 20:80, 40:60).
• Model performance was assessed using five metrics, with optimum sample size defined by 25th percentile performance thresholds.

Main Results:

• For balanced samples (50:50), optimum sample sizes ranged from 750-1,000.
• Unbalanced samples (40:60 ratio) required 1,100-1,300 samples for reliable models.
• Unbalanced samples with a 20:80 ratio consistently failed to produce reliable models.

Conclusions:

• This study provides the first estimates of optimum sample size for Random Forest models at high resolution and extent using simulated data.
• Findings can enhance the reliability of SDMs, optimize field sampling, and improve vector surveillance.
• Further research should validate these findings with real vector data and explore model transferability.