Convolutional neural networks for predicting moisture ratio from images of beetroot cubes under different drying pretreatments

Abstract This study proposes a proof-of-concept approach for predicting the mean moisture ratio of beetroot cubes directly from RGB images acquired during convective drying under different pretreatments and air temperatures. Beetroot ( Beta vulgaris L.) cubes were subjected to four pretreatments: control (BC), ethanol immersion (BE), ultrasound in water (BCU), and combined ethanol + ultrasound (BEU), and dried at 60, 70, and 80 °C. During drying, RGB images of individual cubes were recorded synchronously with gravimetric measurements. A custom convolutional neural network (cnn1v) was developed to perform regression from RGB images to moisture ratio. On the independent test set, the model achieved a global performance of $$R^{2} \approx 0.29$$ , MAE of 0.29, and RMSE of 0.32, indicating moderate overall predictive capability. However, condition-wise analysis revealed substantially higher performance under specific regimes, with $$R^{2} > 0.90$$ and MAE $$< 0.10$$ for BE at 70–80 °C and BCU at 60 °C, while the combined BEU pretreatment exhibited the lowest accuracy across all temperatures. These results demonstrate that the cnn1v model is able to capture meaningful drying trajectories from simple RGB images under controlled conditions, but its performance is strongly dependent on pretreatment and drying regime. The findings highlight both the potential and limitations of low-cost vision-based monitoring for drying processes.