Interpretable Fish Classification through MobileNetV2 and Grad-CAM Visualization

Abstract

Accurate classification of fish species is crucial for monitoring biodiversity and managing fisheries sustainably. This study introduces a deep learning approach leveraging a pre-trained DenseNet201 architecture and transfer learning to classify fish species from images accurately. Trained on over 10,000 images, the model achieved 99.89% accuracy, demonstrating robustness with perfect scores on an extended dataset. Gradient-weighted Class Activation Mapping (Grad-CAM) was employed to confirm that the model focuses on biologically significant features like body shape and fin placement, crucial for accurate identification. These results highlight the model's potential as a reliable tool for automated fish classification, supporting ecological research and sustainable practices in marine environments.