Convolutional neural networks outperform other presence-only species distribution modeling algorithms

Area of Science:

• Ecology
• Biodiversity Modeling
• Conservation Biology

Background:

• Species distribution models (SDMs) are crucial for understanding species occurrences and predicting distributions.
• Most SDMs fail to capture spatial heterogeneity and landscape patterns influencing species, as relationships are hierarchical across spatial extents.
• Convolutional neural networks (CNNs) can process spatial context, including heterogeneity, patterns, and multiscale relationships.

Purpose of the Study:

• To evaluate if CNNs outperform traditional SDM algorithms in predicting species distributions using presence-only data.
• To benchmark CNN performance against widely used algorithms like Maxent and ensemble models.
• To assess the impact of data augmentation on CNN performance, particularly for data-limited species.

Main Methods:

• Modeled 225 species across diverse regions and taxa using presence-only data.
• Compared CNNs against Maxent and ensemble models.
• Assessed the efficiency of data augmentation in CNNs for mitigating sensitivity to limited training data.

Main Results:

• CNNs consistently outperformed other SDM algorithms.
• CNNs with augmented data achieved higher median AUC_ROC (0.77) and AUC_PRG (0.78) compared to ensemble models (0.74 and 0.61).
• For rare species (<30 occurrences), CNNs with augmentation maintained high performance (AUC_ROC = 0.75) exceeding ensemble models (AUC_ROC = 0.68).

Conclusions:

• CNNs effectively incorporate multiscale spatial complexity, enhancing predictive accuracy in species distribution modeling.
• CNNs, especially with data augmentation, offer significant advantages for modeling rare and data-limited species.
• CNNs show potential to revolutionize biodiversity modeling, providing more spatially explicit and ecologically meaningful niche representations for conservation.