One to All: Toward a Unified Model for Counting Cereal Crop Heads Based on Few-Shot Learning.

Area of Science:

• Agricultural Science
• Computer Vision
• Machine Learning

Background:

• Accurate cereal crop head counting is vital for global food security and grain production estimation.
• Current methods lack generalizability, focusing on specific crop varieties and requiring extensive labeling.

Purpose of the Study:

• To develop a unified model, Counting Heads of Cereal Crops Net (CHCNet), for multi-crop head counting using few-shot learning.
• To enhance model generalizability and reduce the cost of data annotation.

Main Methods:

• Employed a refined vision encoder and the Segment Anything Model (SAM) to emphasize crop heads and reduce background noise.
• Introduced a multiscale feature interaction module with a similarity metric for scale-invariant feature learning.
• Utilized a two-stage training procedure: latent feature mining followed by domain-specific feature extraction for inference.

Main Results:

• CHCNet demonstrated superior cross-crop generalization compared to state-of-the-art methods on six diverse datasets (ground and drone imagery).
• Achieved low Mean Absolute Errors (MAEs): 9.96/9.38 for maize, 13.94 for sorghum, 7.94 for rice, and 15.62 for mixed crops.
• The model effectively handles variations in crop head size and shape across different scales.

Conclusions:

• CHCNet provides a robust and generalizable solution for counting multiple cereal crop types with reduced annotation effort.
• The proposed method significantly advances automated crop monitoring and yield prediction capabilities.