Utilizing Geographical Distribution Statistical Data to Improve Zero-Shot Species Recognition.

Area of Science:

• Ecology and Conservation Biology
• Computer Vision and Machine Learning
• Biodiversity Informatics

Background:

• Accurate species recognition is vital for biodiversity conservation and management.
• Traditional methods struggle with data scarcity for rare species and lack generalization.
• Zero-shot learning using Contrastive Language-Image Pre-training (CLIP) shows promise but hasn't incorporated geographic data.

Purpose of the Study:

• To develop a novel CLIP-driven zero-shot species recognition method incorporating geographic distribution knowledge.
• To address the limitations of existing methods in recognizing rare and diverse species.
• To improve the generalization and adaptability of species recognition models.

Main Methods:

• Proposed a CLIP-driven zero-shot species recognition framework.
• Designed prompts utilizing species geographic distribution statistical data.
• Integrated image geocoordinates into species geographic distribution knowledge.
• Calculated similarity using CLIP image and text encoders for recognition.

Main Results:

• Achieved improved zero-shot recognition accuracies across multiple species groups (mammals, mollusks, reptiles, amphibians, birds, insects) on iNaturalist 2021 dataset.
• Demonstrated performance gains ranging from 0.35% to 2.07% compared to standard CLIP prompts.
• Showcased the effectiveness of incorporating geographic distribution data for enhanced species identification.

Conclusions:

• The fusion of geographic distribution data significantly enhances zero-shot species recognition performance.
• This method offers a new approach to leverage species domain knowledge for improved identification.
• The findings provide a valuable contribution to biodiversity monitoring and conservation efforts.