Artificial Intelligence in the Assessment of Seed Physiological Quality: A Critical Review of Spectral, Imaging and Learning-based Approaches for Digital Seed Technology

Abstract

Seed physiological quality, expressed through viability, germination capacity and vigour, determines stand establishment, uniformity and yield potential, yet the reference tests that define it remain destructive, labour-intensive and slow. Over the past decade, artificial intelligence, encompassing classical machine learning and deep learning applied to spectral and image data, has been advanced as a rapid, non-destructive alternative. This review critically evaluates that body of work and asks how far learning-based methods have genuinely improved on established seed testing rather than merely reproduced it under favourable conditions. Evidence was drawn from peer-reviewed studies applying near-infrared spectroscopy, hyperspectral and multispectral imaging, red-green-blue and X-ray imaging, and automated seedling analysis to viability, vigour and germination endpoints across cereal, oilseed, vegetable and tree species. The literature demonstrates consistently high classification and prediction performance under controlled, single-laboratory conditions, with hyperspectral imaging combined with variable selection and convolutional networks emerging as the dominant paradigm for viability and vigour. Deep learning frequently, though not invariably, outperforms chemometric baselines, and transfer learning, data augmentation and generative modelling have been introduced to counter the field's most persistent constraint, namely small and imbalanced datasets. Recurrent weaknesses temper these results: reliance on artificial ageing as a proxy for natural deterioration, narrow cultivar and environmental coverage, scarce external validation, inconsistent reference labelling, and limited interpretability. Reported accuracy therefore reflects internal discrimination more than demonstrated field relevance, and comparability across studies is undermined by heterogeneous endpoints and the absence of shared benchmarks. Progress toward operational digital seed technology will depend less on novel architectures than on representative multi-season datasets, standardised protocols, biologically grounded reference labels and prospective validation against germination and field emergence. The synthesis distinguishes well-supported conclusions from preliminary claims and sets out prioritised research directions to move the field from proof of concept toward dependable, transferable decision support.