Climate-robust evaluation of alfalfa seed maturity via an EMD-guided deep learning framework using multispectral...

Area of Science:

• Agricultural Science
• Computer Science
• Plant Biology

Background:

• Automated plant phenotyping using deep learning struggles with environmental variability.
• Phenotypic plasticity due to annual climatic shifts challenges standard models.
• Accurate alfalfa seed maturity assessment is crucial for crop management.

Purpose of the Study:

• To develop a climate-robust deep learning framework for automated plant phenotyping.
• To address the domain shift problem in multispectral imaging data across different years.
• To reduce annotation costs for phenotyping models in variable agricultural environments.

Main Methods:

• Designed a Multispectral Spatial Attention Network (MSANet) with a 3D-CNN backbone and attention modules.
• Developed an Earth Mover's Distance (EMD)-guided 'diagnose-adapt-finetune' framework.
• Employed EMD-guided Adaptive Batch Normalization (AdaBN) and few-shot fine-tuning.

Main Results:

• MSANet achieved 93% accuracy on single-year data, outperforming baselines.
• Direct model transfer between years resulted in a performance collapse to 41% accuracy.
• The EMD-guided framework restored >90% accuracy on out-of-domain data with minimal samples.
• The adapted model showed resilience to class imbalance and label noise.

Conclusions:

• The proposed framework enhances the generalizability and robustness of AI in phenotyping across diverse agricultural conditions.
• This methodology offers a cost-effective solution for developing reliable AI-driven phenotyping platforms.
• The study highlights the potential of transfer learning and domain adaptation for agricultural applications.