A deep learning-based method for classification, detection, and localization of weeds in turfgrass.

Area of Science:

• Agricultural Engineering
• Computer Science
• Plant Science

Background:

• Precision spraying reduces herbicide use.
• Image classification neural networks can detect weeds in turfgrass.
• Previous methods lacked weed species discrimination and precise location identification.

Purpose of the Study:

• Investigate deep learning models using grid cells for weed detection and localization.
• Evaluate multiple neural networks (DenseNet, EfficientNetV2, ResNet, RegNet, VGGNet) for weed species discrimination and weed/turfgrass classification.

Main Methods:

• Trained deep learning models on grid cells (subimages) to classify presence of weeds.
• Evaluated multi-classifier models for discriminating specific weed species.
• Evaluated two-classifier models for distinguishing weeds from turfgrass.

Main Results:

• VGGNet achieved high F1 scores (≥0.950) for common dandelion and other specific weeds.
• DenseNet, EfficientNetV2, and RegNet showed high performance (≥0.984) for dallisgrass and purple nutsedge detection.
• EfficientNetV2 (two-classifier) demonstrated the highest F1 scores (≥0.981) for weed vs. turfgrass classification.

Conclusions:

• The grid cell-based deep learning method accurately locates weeds in images.
• This approach enables precise, targeted herbicide application via smart sprayers.
• The findings support the integration of machine vision for efficient weed management.