Sample selection bias and presence-only distribution models: implications for background and pseudo-absence data.

Area of Science:

• Ecology
• Computational Biology
• Environmental Science

Background:

• Species distribution models (SDMs) often rely on occurrence data and pseudo-absence (background) data.
• Spatial bias in occurrence data collection (e.g., accessibility) can lead to environmental bias, compromising model accuracy.
• Random background data sampling may not reflect these biases, creating a mismatch.

Purpose of the Study:

• To propose and evaluate a novel method for selecting background data in species distribution modeling.
• To correct for spatial and environmental biases inherent in occurrence data collection.
• To enhance the predictive performance of species distribution models.

Main Methods:

• Investigated the use of 'target-group background' data, sampled with similar spatial bias as occurrence records.
• Compared model performance using target-group background versus randomly sampled background data.
• Evaluated the approach across 226 species, diverse regions, and multiple modeling techniques (GLMs, GAMs, BRTs, Maxent).

Main Results:

• Target-group background data significantly improved average model performance across all considered methods.
• The choice of background data had an impact on predictive performance comparable to the choice of modeling method.
• Performance gains were most substantial when the target-group presence records exhibited strong spatial bias.

Conclusions:

• Using target-group background data is a crucial remedy for spatial bias in species occurrence records.
• This approach offers a practical and effective way to improve the accuracy of species distribution models.
• Increased awareness and application of bias-aware background sampling will advance ecological predictions.