Nondestructive detection of internal quality in multiple peach varieties by Vis/NIR spectroscopy with multi-task CNN method

Abstract

Accurate prediction of the firmness (FI), soluble solids content (SSC), and dry matter content (DMC) of multiple peach varieties is essential for improving market value. However, traditional machine learning methods require training separate models for each quality index, increasing model complexity. To address this, we proposed a multi-task convolutional neural network (MT-CNN) model based on visible/near-infrared (Vis/NIR) spectroscopy. Quality index (FI, SSC, and DMC) and Vis/NIR spectra were measured for four peach varieties (‘Beijing 8’, ‘Baixue’, ‘Shahong’, and ‘Jinchun’), and both single-task (ST) and multi-task models were developed. Important spectral features contributing to the predictions of the MT-CNN models were identified using gradient-weighted class activation mapping (Grad-CAM). The results indicated that both ST-CNN and MT-CNN models achieved good performance. Compared to the ST-CNN model, the MT-CNN model reduced the average RMSE_p for FI and DMC prediction across four peach varieties by 0.0256 N and 0.0719 %, respectively, while the average RMSE_p for SSC increased by only 0.0192 %. Additionally, the global MT-CNN model established based on four peach varieties showed significant improvements, achieving RMSE_p values of 9.4648 N, 0.5904 %, and 0.7628 % for FI, SSC, and DMC, with R values of 0.9496, 0.9615, and 0.9522, and all RPD values exceeding 3. Overall, the proposed MT-CNN model effectively predicted multiple quality of peaches with strong robustness across different varieties.