Effect of Depth Band Replacement on Red, Green and Blue Image for Deep Learning Weed Detection

Area of Science:

• Agricultural Robotics
• Computer Vision
• Machine Learning

Background:

• Automated agricultural robots are increasingly adopted due to lower sensor costs and enhanced computational power.
• Weed detection is established in crops but less developed for pastures, where weeds reduce grazing yield and persist.
• Selective weed cutting is a potential solution, but visual similarity between weeds and forage plants complicates detection with traditional RGB sensors.

Purpose of the Study:

• To investigate the impact of substituting a Red, Green, Blue (RGB) band with depth data on the performance of lightweight YOLOv8 models for weed detection in pastures.
• To evaluate the efficacy of different band combinations, including RGB and RDB (Red, Depth, Blue), for identifying weeds in grazing environments.
• To determine if a depth-enhanced approach can overcome the limitations of RGB-based detection for small, computationally constrained robots.

Main Methods:

• Utilized lightweight YOLOv8 detection models for weed identification.
• Compared the performance of classic RGB (Red, Green, Blue) band combinations against RDB (Red, Depth, Blue) band combinations.
• Evaluated model performance using precision, recall, and mAP50 (mean average precision at 50% intersection over union) metrics.

Main Results:

• The RDB band combination demonstrated superior performance compared to the standard RGB approach for both YOLOv8 small and medium models.
• The RDB combination achieved mAP50 scores of 0.621 for the small model and 0.634 for the medium model.
• The classic RGB approach resulted in lower accuracies, with mAP50 scores of 0.574 and 0.613, respectively.

Conclusions:

• Substituting a Red, Green, Blue (RGB) band with depth data (RDB) significantly improves the detection accuracy of lightweight YOLOv8 models for pasture weeds.
• The RDB approach offers a viable and more accurate alternative to traditional RGB methods, particularly for resource-limited agricultural robots.
• This finding has implications for developing more effective automated weeding systems in pasture management, enhancing grazing yield and sustainability.