Research on Tomato Quality Prediction Models Based on the Coupling of Environmental Factors and Appearance Phenotypes

This study addresses the limitations of current non-destructive techniques for assessing tomato quality, such as their high cost, strong dependence on spectroscopic instruments, and difficulty in dynamic monitoring. The study proposes an integrated tomato quality prediction model that combines a Long Short-Term Memory (LSTM)-based environmental predictor, a Gated Recurrent Unit with attention mechanism (GRU-AT) for dynamic maturity prediction, and a Deep Neural Network (DNN)-based quality evaluation module. The LSTM model demonstrated high accuracy in environmental prediction (R² > 0.9559). The GRU-AT model excelled in color ratio prediction (R² > 0.86), and the DNN model achieved R² values exceeding 0.811 for lycopene (LYC), firmness (FI), and soluble solids content (SSC). Experimental results demonstrate that this approach can accurately predict multiple quality parameters using only standard RGB images. In summary, this study provides a low-cost, low-complexity solution enabling real-time, non-destructive monitoring of greenhouse tomato quality, offering a viable pathway for crop quality management in precision agriculture.