A Physics-Informed Multimodal Pipeline Coupling Convolutional Quality Classification, Q10 Spoilage Kinetics, and a Lightweight Language Model for Smartphone-Based Fruit Shelf-Life Estimation

The waste left over between the production and consumption of fruits and vegetables ranges from 40% to 50%, and a large share of domestic wastage results from people’s inability to assess the shelf life of the produce currently available to them. Machine-vision systems capable of non-destructive food classification have seen rapid development, but they mostly result in a simple label of the product’s freshness, neglecting the storage environment that defines the rate of spoilage and leaving no rationale understandable to a layman. In this paper, the study describes, implements, and characterizes an end-to-end multimodal pipeline that takes a smartphone photograph of a fruit, along with its storage conditions and lightweight additional information supplied by the user, and predicts the number of days remaining to preserve it. The four-class quality classifier (Unripe, Ripe, Overripe, Rotten) was trained using a two-stage transfer learning procedure and a pre-trained ImageNet EfficientNetB3 architecture, processing 7440 curated images of apples, bananas, and oranges from five publicly available datasets. The training corpus size was (7440 images; 6000/960/480 train/val/test split), while the chance-baseline comparison was (25% under four-way balance, i.e., a 3.7-fold improvement). The storage conditions were derived from the freely available weather service, without dedicated hardware, and modeled using a storage mode to reflect actual indoor conditions. The kinetics of postharvest food preservation were incorporated into the model not as a post hoc correction but as part of the engineering features: a Q10-adjusted total shelf life and an elapsed-life fraction, computed with reference storage data. A gradient-boosted regressor accepts the concatenated categorical, environmental, and physics-derived feature vectors and produces the prediction of the days left. The quality propagation between the photographs occurs analytically according to the decay law, not via recurrent neural network classification. An 8B-parameter instruction-tuned language model, complemented by a rule-based deterministic fallback based on reference composition tables, converts numeric predictions into an explanation and storage recommendations. It was implemented using FastAPI and as a cross-platform mobile application. The classifier achieves an accuracy of 92.92% and 0.929 macro-F1 score on a balanced test set (480 images), where the per-class F1 score varies between 0.877 (Overripe) and 0.979 (Unripe), while the remainder of the error falls almost entirely in the vicinity of the Overripe/Rotten classification boundary. The gradient-boosted surrogate model approximates the analytical kinetic model with a mean absolute error of 1.8 days, RMSE of 2.4 days, and R2 = 0.82. Combining visual quality estimation with an explicit kinetic model, rather than training a network to learn the physics of deterioration from images, yields an actionable, interpretable shelf-life prediction for stored products that requires only mobile phone camera sensor input. The study reports the regression metrics explicitly as surrogate-approximation fidelity to an analytical kinetic model, rather than as predictive validation accuracy for stored produce.