A Detection Approach for Wheat Spike Recognition and Counting Based on UAV Images and Improved Faster R-CNN

Area of Science:

• Agricultural Engineering
• Computer Vision
• Remote Sensing

Background:

• Manual wheat spike counting is labor-intensive, inefficient, and prone to inaccuracies.
• Existing automated methods struggle with accuracy, especially in dense canopies and varying field conditions.

Purpose of the Study:

• To develop an innovative unmanned aerial vehicle (UAV)-based intelligent detection method for accurate wheat spike counting.
• To improve the efficiency and accuracy of wheat spike detection and yield prediction in smart agriculture.

Main Methods:

• Collected a high-resolution winter wheat image dataset (2000 images) under diverse fertilization and irrigation conditions.
• Developed an improved Faster R-CNN model using ResNet-50 backbone with residual connections and channel attention.
• Validated the model's performance against manual counts and YOLOv8, and developed a comprehensive yield estimation model.

Main Results:

• Achieved 92.1% mean average precision (mAP) with a 43% reduction in model parameters and improved computational efficiency.
• Demonstrated a 15% reduction in missed detection rate compared to YOLOv8 in dense canopies.
• Spike count regression yielded R² = 0.88, yield prediction errors below 10%, and overall yield estimation accuracy of 93.5%.

Conclusions:

• The proposed UAV-based intelligent detection method significantly enhances wheat spike counting accuracy and efficiency.
• The lightweight and robust model offers practical value for smart agriculture, particularly in yield estimation under optimal management.
• Standardized data acquisition, lightweight model design, and field validation address key challenges in automated agricultural monitoring.