Combining the hyperspectral and image data to predict tobacco starch content using stacking method

Starch content plays a critical role in the quality of tobacco, influencing key attributes such as aroma and sensory quality. Traditional methods for assessing starch content, such as colorimetry and high-performance liquid chromatography (HPLC), have the disadvantages of wasting time and costing expensive. This limits the applicability of measuring starch content in large-scale production. In addition, most existing prediction tasks rely on a single machine learning or deep learning approach, indicating the potential for further enhancement in accuracy. To address this issue, this study developed an ensemble learning model based on the stacking method to improve the prediction accuracy of starch content, using hyperspectral and image data collected from unmanned aerial vehicles (UAVs). Feature selection was performed using synergistic interval partial least squares (SiPLS) on the combined spectral result. A stacking model was developed by integrating support vector regression (SVR) and gated recurrent units (GRU) with multi-layer perceptron (MLP) as the meta-learner. The results indicated that the stacking model achieved high accuacry with a coefficient of determination (R²) of 0.97 and reduced the root mean square error (RMSE) to 1.50 %. Compared to SVR and GRU, the R² values improved by 0.09 and 0.03, respectively, while the RMSE decreased by 2.71 % and 0.53 %, respectively. This methodology significantly enhances prediction accuracy and facilitates precise control over tobacco quality. • A high quality dataset is developed using UAV hyperspectral and visible image features. • An ensemble learning model integrating machine learning and deep learning model. • Non-destructive prediction of tobacco starch is achieved using the ensemble learning model.