Integration of multispectral imaging and explainable AI for alfalfa seed priming optimization under saline-alkali stress

Abstract

Soil salinization poses an increasingly severe threat to sustainable agriculture globally, while conventional seed priming optimization remains time- and resource-intensive, limiting its widespread adoption. This study developed an innovative smart seed priming framework that seamlessly combined multispectral imaging technology with explainable machine learning algorithms to optimize alfalfa seed priming under saline-alkali stress (75 mM Na 2 SO 4 ). Through systematic evaluation of six distinct priming agents (gamma-aminobutyric acid (GABA), ascorbic acid (AsA), melatonin (MT), salicylic acid (SA), spermidine (Spd), and sodium sulfate (Na 2 SO 4 ), our sophisticated stacking ensemble model demonstrated exceptional performance in both priming parameter classification (accuracy: 0.812-0.920, ROC-AUC: 0.941-0.994) and effect prediction (accuracy: 0.820-0.950, ROC-AUC: 0.928-0.960). Comprehensive SHapley Additive exPlanations (SHAP) analysis identified critical spectral features across multiple analytical scales: the parameter model highlighted wavelengths at 570 nm and 970 nm as key indicators of moisture content and flavonoid compositional changes; for priming effects assessment, wavelengths at 850 nm (associated with storage materials) and 490 nm exhibited significant negative synergistic effects specifically on the ‘first’ category of MT priming; detailed individual seed analysis revealed complex nonlinear relationships between priming agent concentration, treatment duration, and corresponding spectral variations. The developed smart seed priming framework enabled rapid and efficient non-destructive optimization of seed priming, significantly accelerating priming technology development processes and advancing sustainable agricultural practices by providing personalized priming solutions under various stress conditions, particularly in regions affected by soil salinization.