Climate-robust evaluation of alfalfa seed maturity via an EMD-guided deep learning framework using multispectral imaging.
- 2026-03
- Plant phenomics (Washington, D.C.) 8(1)
- Zhicheng Jia
- Fang Wang
- Jiayi Fu
- Ruohong Li
- Juan Wang
- Tianqi Zhu
- Shiqiang Zhao
- Chengzhi Lin
- Liru Dou
- Peisheng Mao
- PubMed: 42038795
- DOI: 10.1016/j.plaphe.2026.100184
Study Design
- Methods
- developed a deep learning-based transfer learning framework validated on a multispectral imaging dataset (365-970 nm) covering five maturity stages across three environmentally distinct years
- Funding
- Unclear
Annual climatic and agronomic shifts induce phenotypic plasticity, causing standard deep learning models to fail in high-throughput automated phenotyping tasks, such as alfalfa (Medicago sativa L.) seed maturity assessment. Here, we developed a deep learning-based transfer learning framework to confer climate robustness to such models, validated on a multispectral imaging dataset (365-970 nm) covering five maturity stages across three environmentally distinct years. We designed the Multispectral Spatial Attention Network (MSANet), a hybrid architecture integrating a 3D-CNN backbone with spectral and spatial attention modules to extract complex spatio-spectral features. On single-year data, MSANet achieved 93% classification accuracy, significantly surpassing both traditional Support Vector Machine (77%) and deep learning baselines (e.g., ResNet18, 88%). However, this high intra-year performance did not generalize; direct model transfer to a different year caused accuracy to collapse to 41%, quantifying a profound domain shift. To mitigate this, We proposed an innovative Earth Mover's Distance (EMD)-guided 'diagnose-adapt-finetune' framework. This approach utilized EMD to diagnose layer-specific distributional shifts, employed EMD-guided Adaptive Batch Normalization (AdaBN) to align feature statistics across domains, and concluded with a data-efficient, few-shot fine-tuning strategy. The framework restored predictive accuracy to >90% on out-of-domain data using only 100 labeled samples per class from the target year, representing an approximate 90% reduction in annotation costs compared to full supervision. Crucially, the adapted model exhibited remarkable resilience to real-world data imperfections, maintaining stability under scenarios of class imbalance and label noise. Interpretability analyses further indicated that the model learned biologically plausible spectral correlates associated with seed maturation. Our work presents a generalizable methodology for developing environmentally robust phenotyping platforms, offering a promising pathway to enhance the reliability of AI systems in variable agricultural environments.