Non-Invasive Prediction of Moisture Content of Potato Slices During Drying Using Wavelet Scattering Features and LSTM

This pioneering research explores the application of a wavelet scattering network (WSN) and a long short-term memory (LSTM) model to predict the drying kinetics of potato slices in real-time, non-invasively. This integration captures both localized image textural features using WSN and temporal dynamics via LSTM, thus enhancing the prediction accuracy. Potato slices, 1 mm thick, were dried at 45°C, 50°C, 55°C, and 60°C temperatures. During drying, image texture features were extracted using WSN and used as an input to an LSTM network. Additionally, the quantile score (Q) and prediction-interval-normalised root-mean-square width (PINRW) value were used to evaluate the reliability and robustness of the prediction result. The optimised LSTM network, with 210 hidden neurons, a depth of 3, a dropout rate of 0.40, and a learning rate of 0.0160, achieved an R² of 0.9645 and an RMSE of 0.0649. Uncertainty analysis shows the Q values were low across all drying temperatures: 0.364 (45°C), 0.348 (50°C), 0.398 (55°C), and 0.356 (60°C), indicating high predictive accuracy. Similarly, the PINRW values at α = 95% were 0.132, 0.129, 0.136, and 0.119 for 45°C, 50°C, 55°C, and 60°C, respectively, demonstrating narrow prediction intervals and strong model confidence. This high predictive accuracy allows for reliable, non-invasive, real-time monitoring of moisture content during drying, which has direct implications for industrial drying operations, where improved process control can lead to enhanced product quality, energy savings, and reduced batch rejection, thus resulting in higher throughput, better product consistency and reduced operational costs.