Explainable Deep Learning for Bangladeshi Prawn Species Identification

Aquaculture Shrimp and prawn account for a significant share of the market in Bangladesh, with a 70% share of the agricultural sector. Accounting for 70h is exported in this sector. The species are morphologically indistinguishable, and farmers and exporters report misidentification. Accurate species identification is crucial for promoting species-specific farming, ensuring export quality, and supporting sustainable aquaculture. In this paper, we introduced and evaluated the performance of three DL models: VGG19, ResNet50, and a Custom CNN model for four species (Bagda, Deshi, Golda, Horina) with 6,000 images. We pre-processed and labelled the dataset with augmentation for robustness and evaluated the model using positive and negative precision, recall, and F1-score. We study the thermoregulation of honey bees and utilise SHAP explainability to confirm the models on biologically interpretable features: antennae and body shape. The results showed that the Custom CNN achieved the highest accuracy (97%), followed by VGG19 (96%), and was inferior to ResNet50 (79%). The precision fluctuated slightly in the Custom CNN, but it made the most accurate predictions. Additionally, VGG19 was trained and performed well in prediction. In conclusion, the proposed work results showed that deep learning, specifically VGG19 and a custom CNN, can be effectively explained and is impressively useful practically for prawn species identification, which can be helpful in monitoring and positively impacting the export of shrimp from Bangladesh.