An AI-Driven Dual-Spectral Vision-Language Sensing Framework for Intelligent Agricultural Phenotyping

Area of Science:

• Agricultural Science
• Computer Vision
• Sensor Technology

Background:

• Seed varietal purity and physiological viability are crucial for crop yield and quality.
• Non-destructive seed assessment faces challenges in variety discrimination and internal defect detection.

Purpose of the Study:

• To develop an AI-driven multimodal sensing framework (S3-Net) for autonomous seed quality evaluation.
• To enhance fine-grained variety discrimination and internal defect perception in seeds.

Main Methods:

• Integration of vision-language alignment (Knowledge-Vision Alignment module) with dual-spectral sensor fusion (Dual-Spectral Fusion module).
• Utilizing encyclopedic morphological descriptions for feature learning and few-shot generalization.
• Combining high-resolution RGB and Short-Wave Infrared (SWIR) sensing for external and internal trait characterization.

Main Results:

• S3-Net achieved 96.9% accuracy for species identification and 95.8% for viability detection on a dataset of 6000 samples across 12 crop categories.
• Outperformed ResNet-50 by 40.3% in extreme 1-shot learning scenarios.
• Demonstrated a stable inference throughput of 95 frames per second, suitable for industrial applications.

Conclusions:

• S3-Net provides a robust and efficient solution for intelligent agricultural phenotyping.
• The multimodal approach significantly improves seed quality evaluation accuracy and generalization capabilities.
• The framework addresses limitations in current non-destructive seed assessment methods.