Accurate fine-grained weed instance segmentation amidst dense crop canopies using CPD-WeedNet.

Area of Science:

• Agricultural technology
• Computer vision
• Machine learning

Background:

• Accurate farmland weed segmentation is crucial for targeted weeding and sustainable agriculture.
• Challenges include similar crop/weed morphology, occlusions, lighting variations, and scale diversity, hindering current methods on resource-constrained platforms.

Purpose of the Study:

• To propose a novel instance segmentation framework, CPD-WeedNet, for fine-grained weed identification in complex agricultural environments.
• To enhance accuracy and efficiency for real-time intelligent weeding systems.

Main Methods:

• Developed CPD-WeedNet with three core components: CSP-MUIB backbone for enhanced feature discrimination, PFA neck for integrating shallow details, and DFS neck utilizing Transformer for global context.
• Evaluated on a self-constructed soybean field weed dataset and the public Fine24 dataset.

Main Results:

• CPD-WeedNet achieved 80.6% mAP50(Mask) and 85.3% mAP50(Box) on the soybean dataset, outperforming YOLO baselines.
• On the Fine24 dataset, it reached 75.4% mIoU and 81.7% mAcc, showing a balance between performance and efficiency.
• Pixel-level metrics (mIoU, mAcc) reached 86.6% and 94.6% on the soybean dataset.

Conclusions:

• CPD-WeedNet demonstrates significant potential for low-cost, real-time intelligent weeding systems.
• The framework offers a robust solution for precise weed segmentation in challenging agricultural conditions.
• This research advances precision agriculture through improved weed identification technology.