A Novel Oyster Freshness Prediction Model Based on Hierarchical Feature Fusion With XGBoost and CNN-LSTM for Edge Deployment

Oysters, as highly perishable seafood, undergo rapid quality deterioration influenced by storage temperature and duration. This paper presents a hybrid XGBoost-CNN-LSTM model for real-time quality assessment, which synergistically integrates electronic nose data with multimodal freshness indicators-sensory scores, Aerobic Plate Count (APC), and Total Volatile Basic Nitrogen (TVB-N). The proposed architecture employs a dual-path feature fusion strategy, combining raw sensor inputs with XGBoost-optimized features to construct a discriminative hierarchical feature matrix. A CNN layer extracts local patterns from the fused features, followed by a single-layer unidirectional LSTM with 32 hidden units to capture temporal dependencies. The decoder efficiently maps final hidden states through fully connected layers to simultaneously predict sensory, microbiological (APC), and chemical (TVB-N) freshness indicators. Comparative evaluations against two baseline models (TCN and SqueezeNet) demonstrate the superior performance of the proposed model in both prediction accuracy and computational efficiency. Validated on embedded platforms (Raspberry Pi 4B and Jetson TX2 NX), the model achieves high inference speed with minimal resource consumption, establishing a practical and efficient framework for intelligent seafood quality monitoring in edge environments.