Explainable Multi-task Convolutional Neural Network Framework-Assisted Simultaneous Prediction of Cherry Maturity-related Traits Using Vis-NIR Spectroscopy

Accurate and non-destructive assessment of cherry maturity requires the simultaneous quantification of multiple correlated quality traits, including firmness, soluble solid content (SSC), color ( L*, a*, b* ), titratable acidity (TA), and vitamin C. Visible-near-infrared (Vis-NIR) spectroscopy offers a rapid acquisition of such information; however, existing modeling methods require laborious preprocessing and feature-wavelength selection, while also necessitating the establishment of separate predictive models for different traits. To address this, We propose a multi-task deep learning framework WOA-MHA-MCNN integrated with post-hoc explanation approach (shapley additive explanations, SHAP) for enhanced interpretability, which optimizes a multi-task convolutional neural network model by incorporating whale optimization algorithm (WOA) and multi-head attention (MHA). The proposed model achieves excellent performance in multi-trait prediction (0.954 < R v 2 < 0.975, 5.09 < RPD < 6.59). Additionally, SHAP are employed to attribute wavelength contributions for each trait, revealing key spectral regions driving the predictions and enhancing model transparency. This study provides an efficient solution for the rapid and accurate evaluation of multiple maturity-related internal quality parameters in cherries using Vis-NIR spectroscopy.