Stacked Deep Learning-Based Quality Classification of Dried Fish Products for Post Harvest Loss Reduction and Value Chain Enhancement

This study addresses critical inefficiencies in traditional fish drying practices by developing an artificial intelligence (AI)-powered classification system to standardize quality control and enhance market competitiveness for coastal communities. Unlike existing works that rely solely on isolated classifiers, this study introduces a hybrid stacked architecture incorporating both deep and traditional machine learning models, enabling improved interpretability and robustness. Using a threestage computational methodology, we integrated deep learning with feature optimization to overcome the longstanding challenges of manual quality assessment. The framework consists of: (1) the utilization of a publicly available dataset containing 8,290 high-resolution images of five commercially important species under controlled drying conditions, (2) hierarchical feature extraction using ResNet50 with Lasso regression (LR) for dimensionality reduction, and (3) a novel ensemble learning strategy based on stacked generalization, combining four diverse classifiers trained on optimized deep feature representations. Unlike prior studies that use off-the-shelf convolutional neural network (CNN), our model integrates LR atop CNN embeddings before stacking, thereby redefining classifier synergy in the dried fish domain. The proposed model achieved 99.94% classification accuracy, outperforming conventional single-algorithm approaches by 0.06%, while reducing computational overhead by 34% through optimized feature selection. This study introduces a novel integration of deep features and ensemble stacking methods for dried seafood authentication. This novel design effectively addresses critical texture and color variability issues arising from inconsistent drying processes. The non-destructive nature of this approach enables real-time species identification with a 0.06% misclassification risk, aligning with FAO priorities to minimize postharvest losses in small-scale fisheries. Furthermore, this study offers a replicable AI blueprint for other agricultural commodities, demonstrating methodological scalability and domain adaptability. By integrating ancestral preservation techniques with Industry 4.0 innovations, this framework enhances processing efficiency, ensures product traceability, and promotes equitable income distribution for artisanal producers.