Early sweet potato black spot prediction via multimodal data fusion

This is the authors' abstract. We don't add key points for this paper.

The main pathogen of sweet potato black rot, Ceratocystis fimbriata , induces the production of toxic secondary metabolites, leading to significant post-harvest economic losses. Establishing early rapid detection technologies is crucial for ensuring sweet potato food safety and reducing economic losses. This study systematically monitored the spectral image (HSI), electronic nose (E-nose) response signal, and total phenolic content (TPC) reference value of sweet potato samples after artificial inoculation with the pathogen, aiming to utilize TPC as a key biochemical indicator for the early prediction of disease progression. The experiment compared single-source and multi-source data fusion methods. Results showed that the CARS-PCA-MHA-CNN model(Parameters was reduced by 96.58 %) based on a feature-level fusion strategy achieved the best predictive performance (R²=0.974, RMSEP=0.041, RPD=6.14). Compared with single-source data, the prediction accuracy was improved by 7.6 % and 6.4 %, respectively. Furthermore, the model's generalization ability was tested on an independent test set (unenhanced). This study proposes a reliable and non-destructive method for the early prediction of postharvest diseases in root and tuber crops, which has great application potential in the field of intelligent monitoring of agricultural products. • Changes in TPC are a key indicator of pathological changes in sweet potatoes. • Complementary data information improves model robustness. • Multi-head attention mechanism was used to address data scale heterogeneity. • The generalization ability of the MHA-CNN model was validated.