A deep learning-based fusion framework for robust fine-grained classification of sea turtles in support of marine biodiversity

Abstract

Accurate classification of sea turtle species is crucial for ecological monitoring and conservation, yet traditional visual classification methods remain limited by underwater imaging challenges such as occlusions, poor lighting, and background noise. To address these limitations, we propose an enhanced deep learning-based classification framework that integrates both color and structural features to improve the robustness of species recognition in complex marine environments. Building upon the ResNet-50 backbone, we introduce a four-channel input tensor comprising RGB data and Sobel-filtered edge maps, capturing both semantic and morphological information. Two novel fusion modules, LiteAFNet and AlphaBlendNet, are designed to integrate these features effectively. LiteAFNet leverages a lightweight attention mechanism to highlight discriminative regions, while AlphaBlendNet adaptively balances RGB and edge cues based on spatial context. Experimental results demonstrate significant improvements in classification performance across all evaluation metrics. Specifically, AlphaBlendNet achieves the highest precision (0.84), recall (0.88), F1-score (0.86), and mean average precision (mAP) of 87.2%, outperforming both the baseline fusion and LiteAFNet configurations. These results indicate that integrating color histograms with structural edge features enhances the model’s ability to distinguish between species with similar visual traits. This framework offers a scalable, accurate, and automated solution for underwater species classification and holds potential for broader application in marine biodiversity monitoring.