Intelligent systemic microbial monitoring method for fresh food supply chain safety

Introduction Microbial safety monitoring in fresh-food systems increasingly relies on heterogeneous information sources, including hyperspectral measurements, microscopy images, structured environmental variables, and temporally related observations. These information sources are not uniformly available across monitoring scenarios, making it inappropriate to impose the same temporal, spectral, or contextual assumptions on every sample. This study proposes the ForwardSpectralProbabilistic (FSP) framework for microbial classification and pathogen-risk prediction under heterogeneous sensing conditions. Methods FSP employs modality-specific encoders to map spectral, image, and structured observations to a common latent dimensionality while preserving their native input representations. Contextual information is incorporated through gated residual fusion when environmental, operational, or acquisition metadata are available. The Forward Consistency Projector generates temporally aligned latent evidence only for observations with a valid predecessor target correspondence, whereas the Spectral Consistency Mapper operates only on measurements with a native spectral axis. The Probabilistic Mixture Integrator then combines the available current, temporal, and spectral representations through availability-normalized learned weights. Harmonic State Projection (HSP) is further introduced as a structured regularization strategy that constrains temporal projection consistency and preserves relationships among low-frequency spectral components without constructing unavailable modalities. The framework is evaluated independently on four publicly accessible microbial-safety datasets covering Salmonella serovar classification from hyperspectral measurements, microscopy-based bacterial classification, poultry Salmonella-Campylobacter risk prediction, and agricultural-water microbial marker prediction. The experiments use modality-matched comparison methods, grouped or stratified data partitioning, and repeated random seeds. Results and discussion The descriptive results show that FSP is competitive across heterogeneous tasks, while its strongest observed performance is not obtained uniformly for every metric and dataset. Spectral-consistency and harmonic-relation modeling provide the clearest gains for native hyperspectral measurements, meteorological context and temporally aligned evidence contribute to the structured poultry-risk task, and acquisition metadata improve robustness under microscopy domain shift. Modality-aware ablation experiments further indicate that spectral, temporal, and contextual mechanisms provide complementary contributions only when the corresponding information is available. FSP provides an availability-aware formulation for microbial monitoring that maintains a consistent prediction framework without imposing artificial spectral or temporal structure on heterogeneous food-safety data.