Non-destructive detection of microplastics stress in rice seedling: an interpretable deep learning approach using excitation emission matrix fluorescence spectra of root exudates

Introduction: Microplastics (MPs), ubiquitous and insidious pollutants pervading agricultural systems, pose an escalating threat to global food security. This makes the development of nondestructive methods for the early detection of MPs stress in rice seedling an urgent scientific imperative. Method: Rice seedlings were cultivated under exposure to polyethylene terephthalate (PET), polystyrene (PS), and polyvinyl chloride (PVC) MPs at concentrations of 0 (control), 10, and 100 mg/L. Based on the stress-induced alterations in root exudates composition, a novel detection method for MPs stress in rice seedlings was developed using excitation-emission matrix fluorescence (EEMF) spectra combined with deep learning. Results: Analysis of the original EEMF spectra revealed discernible differences. Feature extraction was performed using both the peak method and the PARAFAC method. Spectral changes in seedlings exposed to the low MP concentration (10 mg/L) were relatively minor compared to the control group. In contrast, exposure to the high concentration (100 mg/L) induced significant alterations in humic acid-like and amino acid-like substances. Subsequently, enhanced Vision Transformer (VIT) models were developed utilizing three distinct data representations: full EEMF spectra, emission spectra at specific excitation wavelengths, and extracted characteristic fluorescence values. The optimal model achieved 100% classification accuracy. Furthermore, SHapley Additive exPlanations (SHAP) analysis was employed to evaluate feature importance, identifying both humic acid-like and marine humic acid-like components as major contributors to the model’s predictions. Conclusion: In summary, this study establishes a novel, non-destructive, and interpretable framework for the early detection of MPs stress in rice seedlings based on EEMF spectra of root exudates combined with deep learning.