Adaptive multi-scale feature refinement for wheat phenology recognition using cross-scale attention mechanisms.

Area of Science:

• Agricultural Science
• Computer Science
• Plant Science

Background:

• Accurate crop phenotyping is crucial for agricultural management but challenging due to subtle growth transitions and environmental noise.
• Existing computational methods struggle with fine-grained wheat stage identification from ground-level imagery.

Purpose of the Study:

• To introduce AMFR-Net, an Adaptive Multi-Scale Feature Refinement Network, for precise wheat phenotyping.
• To enhance the identification of wheat growth stages using RGB imagery by addressing limitations of conventional architectures.

Main Methods:

• Developed AMFR-Net, incorporating a ResNet-101 backbone and a novel Adaptive Multi-Scale Attention Fusion (AMSAF) module.
• AMSAF utilizes cross-scale interaction blocks and confidence-weighted feature aggregation for hierarchical recalibration of spatial-semantic representations.
• The network adaptively amplifies phenologically relevant features and suppresses irrelevant context for robust generalization.

Main Results:

• AMFR-Net achieved state-of-the-art performance on the CGIAR benchmark, with Top-1 Accuracy of 89.10%, Macro-F1 of 89.10%, and AUC of 97.88%.
• Demonstrated superior discriminability between phenologically adjacent wheat stages compared to baseline CNN models.
• Ablation studies confirmed the effectiveness of multi-level attention and scale-aware refinement.

Conclusions:

• AMFR-Net provides a scalable, interpretable, and field-deployable solution for in-situ wheat phenology monitoring.
• The proposed framework improves fine-grained stage identification accuracy and robustness.
• Establishes a foundation for integrating multimodal sensing, weak supervision, and cross-seasonal adaptation in phenotyping.