WPDSI: A deep learning method for wheat phenology detection from single-temporal images

Area of Science:

• Agricultural Science
• Computer Science
• Remote Sensing

Background:

• Accurate wheat phenology monitoring is vital for global food security.
• Deep learning models using multi-temporal images improve wheat phenology detection but face challenges like complexity and real-time deployment.
• Current methods struggle with computational efficiency and practical field application.

Purpose of the Study:

• To develop an optimized method for deriving wheat phenology from single-temporal images (WPDSI).
• To reduce model complexity and data requirements for efficient wheat phenology detection.
• To enhance the accuracy and real-time applicability of automated wheat phenology monitoring.

Main Methods:

• Knowledge distillation: A teacher model trained on multi-temporal images guides a student model using single-temporal images.
• Multi-layer attention transfer: Enables the student model to learn features from multiple layers of the teacher model.
• Development of the Wheat Phenology from Single-temporal Images (WPDSI) model.

Main Results:

• The WPDSI method achieved an overall accuracy (OA) of 0.927, comparable to multi-temporal models.
• Demonstrated strong generalization capabilities on unseen datasets.
• Significantly improved real-time performance and computational efficiency.

Conclusions:

• The proposed WPDSI method offers a practical and efficient solution for field-based wheat phenology derivation.
• Combines high accuracy with reduced complexity and enhanced real-time capabilities.
• Facilitates more accessible and scalable automated monitoring of wheat growth stages.