Deep learning-based image classification of turtles imported into Korea.

Area of Science:

• Ecology
• Computer Science
• Zoology

Background:

• Turtles are vital for ecosystem health but face endangerment due to global trade.
• Imported non-native turtles pose risks as invasive alien species, necessitating control measures.
• Rapid and accurate classification systems are essential for turtle conservation and invasive species detection.

Purpose of the Study:

• To develop and evaluate a deep learning model for classifying 36 imported turtle species.
• To identify the most effective model for rapid and accurate identification of turtles in Korea.
• To support the management of turtle trade and the detection of invasive alien species.

Main Methods:

• Eight Single Shot MultiBox Detector (SSD) models with different backbone networks were employed.
• Turtle images were sourced from Google, identified morphologically, and divided into training, validation, and test sets.
• Data augmentation was utilized to enhance model robustness and prevent overfitting.

Main Results:

• The Resnet18 model achieved the highest mean Average Precision (mAP) of 88.1%.
• Resnet18 demonstrated the fastest inference time at 0.024 seconds.
• The average correct classification rate for 36 turtle species using Resnet18 was 82.8%.

Conclusions:

• The Resnet18 model shows significant potential for accurate and efficient classification of imported turtles.
• This technology can aid in managing the turtle trade and detecting alien invasive species in natural environments.
• The study provides a valuable tool for ecological conservation efforts and biosecurity.