Physics-Informed Sequential Learning with KAN-Mamba for Quality Prediction in Agricultural Cold-Chain Data

Abstract

Fresh-cut vegetables are easily damaged by enzymatic browning after cutting because cell compartmentalisation is disrupted, oxygen diffusion accelerates, and moisture migration changes at the exposed tissue surface. Develop a physics-informed KAN-Mamba model for browning prediction in fresh-cut lettuce, potato, lotus root and cucumber under cold-storage conditions in this study. The model uses Kolmogorov-Arnold network layers for nonlinear multimodal feature mixing, selective Mamba state-space blocks for long-term storage memory, and a soft reaction-consistency constraint based on temperature-sensitive oxidation and browning-rate monotonicity. The input features are RGB colour indices, hyperspectral band ratios, headspace oxygen, relative humidity, storage temperature, vegetable identity and elapsed storage time. In a 72-hour experiment with 1,920 time-stamped records, the proposed model achieved an RMSE of 3.18, an MAE of 2.41, a freshness-grade accuracy of 93.6%, and a high-browning warning recall of 91.7%, and outperformed the conventional machine-learning and sequence-model baselines.