Identification of sea turtles using adaptive graphs and attention networks: a comparative analysis of RAG, KNN, and ALSG-based methods

Graph-based algorithms provide an innovative framework for modeling morphological features in sea turtles, such as scute geometry, color patterns, and unique shell and flipper structures. These models represent individual images through graphs, where nodes encode anatomical attributes and edges define spatial or functional relationships. This study evaluates three graph construction methods: Region Adjacency Graph (RAG), the adjoining continuity in space is preponderant; KNearest Neighbors Graph (KNN), the neighborhood relationship establishes your strength; and the Adaptive Local Similarity Graph (ALSG), designed to dynamically adjust to morphological variations. Each method is integrated into an optimized Graph Attention Networks (GAT) architecture, where attention mechanisms enable nodes to focus on relevant neighborhood information, enhancing performance in uncontrolled scenarios. The effectiveness of each method depends on graph structure: RAG preserves anatomical topology, KNN captures global similarities, and ALSG identifies adaptive local patterns, crucial for handling variations in lighting, angles, or obstructions. This analysis shows that method selection directly impacts individual identification robustness, particularly when processing heterogeneous data typical of in-field monitoring. The synergy between adaptive graphs and GATs eliminates the need for controlled environments, enabling practical sea turtle conservation applications. For instance, precise identification under variable conditions supports migration tracking, population health assessment, and threat detection without physical intervention. This work not only improves photoidentification but also enriches ecological studies by generating detailed databases on behavior and sea turtle population dynamics. The rigorous comparison highlights that KNN3 combined with GATs, outperforms RAG and ALSG in accuracy and adaptability by leveraging optimal neighborhood relationships to capture global similarities. This approach promotes sustainable conservation strategies by reducing costs and increasing scalability, aligning computational tools with urgent marine preservation needs. Thus, innovation in graphical representations becomes a cornerstone for effective biodiversity management in dynamic and complex environments.