Computer vision for assessing species color pattern variation from web-based community science images.

Area of Science:

• Ecology
• Computer Science
• Bioinformatics

Background:

• Community science digital vouchers offer vast phenotypic data but require extensive manual effort for extraction.
• Automating phenotypic data extraction is crucial for large-scale ecological studies.

Purpose of the Study:

• To develop and validate a computer vision model for automatic categorization of species color patterns.
• To document the striped/unstriped color polymorphism in the Eastern Red-backed Salamander (Plethodon cinereus) using community science images.
• To analyze niche overlap and breadth between color morphs at a range-wide scale.

Main Methods:

• An ensemble convolutional neural network model was employed to analyze 20,318 iNaturalist images.
• The model was trained to categorize color patterns, specifically the striped/unstriped polymorphism.
• Image data from community science platforms were utilized.

Main Results:

• The computer vision model achieved high accuracy (approximately 98%) in categorizing color patterns, even with heterogeneous image data.
• Extensive niche overlap was identified between the striped and unstriped morphs.
• Striped morphs exhibited a wider niche breadth across the species' range.

Conclusions:

• Machine learning, particularly computer vision, can effectively process heterogeneous community science image data for ecological research.
• The developed workflow enables large-scale analysis of phenotypic variation and its ecological implications.
• This approach significantly reduces human effort in data extraction from digital vouchers.