Alfalfa (Medicago sativa L.) flowers indeterminately, so seeds on the same plant mature asynchronously and harvested lots mix immature, mature, and overmature seeds. Traditional maturity tests are destructive, labor-intensive, and subjective, and they cannot meet modern agriculture’s demand for automated, standardized, high-throughput screening, while multispectral imaging offers a non-destructive alternative. Deep learning can classify spectral images, but most models are trained on single-year data and fail when climate, soil, and irrigation vary. Inter-annual “physiological drift” changes seed biochemistry and spectral signatures, and conventional transfer learning lacks diagnostics, relying on trial-and-error adaptation. Given these challenges, deeper research is needed to develop climate-robust, transferable, and data-efficient models for alfalfa seed maturity assessment.

Researchers at the College of Grassland Science and Technology, China Agricultural University, Beijing, China, developed MSANet and an Earth Mover’s Distance (EMD)-guided domain adaptation framework for alfalfa seed maturity. The study was published (DOI: 10.1016/j.plaph.2026.100184) in Plant Phenomics (2026, Volume 8, Article 100184). It analyzed 29,912 multispectral seed images collected across 2022, 2023, and 2024 and five developmental stages, testing whether a model trained in one year could be adapted to another year with minimal labeling, and evaluated cross-year transfer under real-world imperfections.

The Multispectral Spatial Attention Network (MSANet) first applies spectral attention through global and local branches, then extracts spatio-spectral features with a 3D-CNN, and finally models long-range spatial dependencies using an eight-head Transformer spatial processor. On single-year data, MSANet reached 93% classification accuracy, outperforming a support vector machine (SVM; 77%) and ResNet18 (88%). However, direct transfer to a different year dropped accuracy to 41–50%, exposing a large domain shift. The EMD-guided framework diagnosed layer-specific distribution gaps, selectively adapted batch normalization (AdaBN) statistics, and fine-tuned only a small labeled target set. With 100 samples per class, accuracy returned to more than 90%, cutting annotation costs by about 90% versus full supervision. The model remained stable under class imbalance and label noise. Interpretability analyses showed progressive class separation and gradient-weighted class activation mapping (Grad-CAM) attention shifting from local regions in immature seeds to the whole seed surface at maturity. Spectral signals correlated negatively with chlorophyll and positively with crude protein (CP), supporting biologically plausible learning.

The authors said the advance is less about one high score and more about diagnosing why models break across years. They said that instead of retraining from scratch, researchers can use EMD as a distributional ruler to find the layers most affected by climate-driven shifts. They explained that EMD-guided AdaBN aligns those layers, and few-shot fine-tuning then corrects decision boundaries with minimal labels. They added that this makes climate-robust seed phenotyping practical for breeding and quality testing, while cautioning that broader locations, cultivars, and extreme weather still need validation.

For the seed industry, the framework could support more consistent harvest timing, more uniform seed lots, and faster non-destructive quality checks even when growing seasons differ. Because it needs only about 100 labeled seeds per maturity class, it could reduce the cost of building year-specific models and make artificial intelligence (AI)-based phenotyping more feasible in variable agricultural environments. The same diagnose-adapt-finetune logic may be extended to other crops and spectral imaging tasks where climate, location, or management practices create domain shifts. Future work should integrate meteorological data such as growing degree days, test unsupervised domain adaptation, and validate performance under extreme weather events.

###

References

DOI

10.1016/j.plaphe.2026.100184

Original Source URL

https://doi.org/10.1016/j.plaphe.2026.100184

Funding information

This research was supported by the earmarked fund for CARS (CARS-34), and Research of key technology for seed production in alfalfa with salinity-alkalinity and drought stress (2022ZD0401105).

About Plant Phenomics

Plant Phenomics publishes breakthrough research that integrates genomics, genetics, physiology, molecular biology, bioinformatics, statistics, mathematics and computer science to advance plant phenotyping—from the cellular level to entire populations. The journal tackles key scientific challenges in phenomics by showcasing cuttingedge technologies for data acquisition, management, interpretation and modelling. Its scope spans practical applications across all plant sciences, aiming to translate novel phenotyping tools into realworld solutions for breeding, crop management and fundamental plant biology. By fostering interdisciplinary innovation, Plant Phenomics serves as a hub for researchers developing highthroughput, accurate and scalable methods to decipher complex plant traits, ultimately accelerating progress in agriculture and plant science.