Deep learning-based image classification of sea turtles using object detection and instance segmentation models.

Area of Science:

• Marine Biology
• Computer Science
• Artificial Intelligence

Background:

• Sea turtles are highly migratory and inhabit diverse environments, making population monitoring difficult.
• Deep learning (DL) models, particularly object detection, show promise for wildlife monitoring but struggle with complex backgrounds.
• Instance segmentation models offer improved accuracy for complex image classification compared to traditional object detection.

Purpose of the Study:

• To compare the performance of YOLOv5 (object detection) and YOLOv5-seg (instance segmentation) for detecting and classifying sea turtles.
• To evaluate the effectiveness of these DL models in challenging, complex image datasets.

Main Methods:

• Utilized image datasets from iNaturalist and Google, divided into training (64%), validation (16%), and testing (20%) sets.
• Employed YOLOv5 for object detection and YOLOv5-seg for instance segmentation to identify sea turtles.
• Assessed model performance using loss functions and mean Average Precision (mAP) metrics.

Main Results:

• YOLOv5-seg exhibited a lower error rate in detection compared to YOLOv5, based on loss functions.
• The YOLOv5-seg model demonstrated superior performance with higher mAP values (mAP0.5: 0.918, mAP0.5:0.95: 0.831) than YOLOv5 (mAP0.5: 0.885, mAP0.5:0.95: 0.795).
• Instance segmentation (YOLOv5-seg) proved more effective for both detecting and classifying sea turtles in complex imagery.

Conclusions:

• YOLOv5-seg offers enhanced accuracy for sea turtle detection and classification, especially in challenging underwater or complex background images.
• This study highlights the potential of instance segmentation models to significantly improve wildlife monitoring efforts.
• The findings can contribute to more robust and efficient sea turtle conservation and management strategies.