Application of machine learning models for non-invasive seed quality detection

The integration of artificial intelligence (AI) in agriculture has revolutionized traditional farming practices, addressing challenges in food security, sustainability, and climate change. In seed science, AI-driven models enhance seed quality assessment, moving beyond conventional time-consuming and invasive methods. This study presents a pipeline that combines deep learning and machine learning approaches to predict legume seed germination potential using heterogeneous features, including color, physical traits, and chemiluminescence data (ultra-weak photon emission and delayed luminescence). A dataset of 1,038 seed samples from five legume species was analyzed, aimed at finding the most informative features to discriminate germination potential, and evaluating whether classification performance could reach promising levels. Results demonstrated that machine learning models trained using color and physical features outperform those relying only on chemiluminescence data. Notably, the best-performing model leveraged gradient boosting techniques and reached about 80% prediction accuracy. Our findings underscore the importance of multimodal approaches in seed quality assessment, highlighting the role of AI in advancing non-invasive agricultural diagnostics. This research contributes to precision agriculture by providing a promising AI-powered framework for seed quality evaluation, based on selected features, which could potentially support enhanced crop yield and sustainability. • A 1.038-entry 5-species dataset with multimodal features was used to build classifiers. • Color and physical traits were predictive features of germination. • The best models included UPE/DL and color/physical data, or color/physical data only. • 80% accuracy in germination capability prediction was obtained. • Prediction was done without using information on species or storage condition.