<![CDATA[Turtles are remarkable organisms as they play a vital role in maintaining a balanced and healthy ecosystem. However, their existence is at risk due to growing threats of habitat loss, illegal trade and environmental degradation. To support conservation and monitoring efforts for these turtle species, this study proposes TurtleNet, a custom deep learning model designed to accurately identify turtle species with acute precision. The proposed model relies on a customized EfficientNetB4 architecture as a backbone with an enhanced lightweight attention mechanism of Squeeze and Excitation blocks, allowing it to focus on subtle visual features that distinguish closely related species. The model was trained and evaluated on our curated Turtle1 dataset as well as the images.cv turtle dataset. The performance comparison of TurtleNet with 11 individual state-of-the-art deep learning classification models showed its superior ability by classifying Asian turtle species with 97.25% accuracy. To further enhance classification performance, TurtleNet is integrated with the YOLOv8 model using a soft-voting ensemble strategy, thereby increasing the classification accuracy by 2.41%. The results demonstrate that TurtleNet has strong potential for automated species monitoring and conservation using non-invasive measures. Grad-CAM visualization is used to highlight discriminative morphological regions relevant to image identification and is used to interpret the feature selection process of the proposed model.]]>
Copyrights © 2026
