Deep Learning Model with Attention Mechanism for Biometric Re-Identification of Green Sea Turtles in Long-Term Tracking Scenario
GPS tags, plastic tags, and Inconel tags that are difficult to apply and degrade over time are currently used for re-identification of Chelonia mydas sea turtles. A deep learning model with attention mechanism is proposed for biometric re-identification of individual sea turtles in a long-term scena...
Saved in:
| Main Author: | |
|---|---|
| Format: | Thesis |
| Language: | English |
| Published: |
2024
|
| Subjects: | |
| Online Access: | http://ir.unimas.my/id/eprint/46057/1/Thesis%20Master_Khalif%20Amir%20Bin%20Zakry.pdf |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | GPS tags, plastic tags, and Inconel tags that are difficult to apply and degrade over time are currently used for re-identification of Chelonia mydas sea turtles. A deep learning model with attention mechanism is proposed for biometric re-identification of individual sea turtles in a long-term scenario. The deep learning model with attention mechanisms re-identifies sea turtles on which it has been trained via its carapace. The carapace pattern is a stable feature that would change over time due to scratching or barnacle growth, making re-identification a non-trivial task. Results of testing with datasets obtained from sampling efforts at Talang Besar, Sarawak, using a deep learning model with attention mechanism achieves 97.23% accuracy on direct re-identification and 73.61% accuracy on long-term re-identification scenario. This is comparable to an AlexNet CNN model in direct re-identification and showed 1.39% accuracy improvement in long-term scenario tests. The attention-based model however is -1.38% and -2.77% less accurate than SqueezeNet and ResNet50 CNN models in direct re-identification and inaccurate by -6.95% and -18.06% respectively than SqueezeNet and ResNet50 CNN models in long-term scenario tests. The work shows that deep learning model with attention mechanisms can perform re-identification directly and in a long-term simulated scenario of barnacle growth with reliable accuracy above 70% but still underperforms when compared to some CNN models trained and tested on the same tasks. The findings of this work show that machine learning models can be used for the re-identification of Chelonia mydas sea turtles and that such models can be adopted for further use animal re-identification purposes. |
|---|