Semantic segmentation for weed detection in corn.

Area of Science:

• Agricultural Science
• Computer Vision
• Machine Learning

Background:

• Weed detection is vital for precision agriculture but challenging due to diverse weed species and growth stages.
• Current deep learning models require extensive datasets, which are impractical to create for all weed variations.
• A novel approach is needed to simplify weed detection and improve management strategies.

Purpose of the Study:

• To develop an efficient and accurate weed detection method for precision agriculture.
• To simplify the weed detection process by focusing on crop segmentation rather than direct weed identification.
• To reduce the complexity of training data requirements for weed detection models.

Main Methods:

• Utilized semantic segmentation to create a precise mask of crop pixels (e.g., corn).
• Identified all green vegetation outside the crop mask as weeds (indirect detection).
• Optimized the semantic segmentation model using knowledge distillation for improved real-time performance.

Main Results:

• The optimized DeepLabV3+ model achieved over 99.5% average accuracy (aAcc).
• Mean intersection over union (mIoU) across all categories exceeded 95.5%.
• The model demonstrated a processing speed of over 34 frames per second (FPS).

Conclusions:

• A novel method accurately segments crops, enabling indirect weed identification outside the crop mask.
• This approach bypasses the challenges of diverse weed species, densities, and growth stages.
• The method provides a practical and efficient solution for training computer vision models for precision weed management.