Deep learning-based weed detection for precision herbicide application in turf

Area of Science:

• Agricultural Science
• Computer Science
• Robotics

Background:

• Precision weed mapping is crucial for efficient herbicide application in turf management.
• Smart sprayers require accurate weed identification and susceptibility data for targeted spraying.
• Deep convolutional neural networks (DCNNs) offer potential for advanced weed mapping.

Purpose of the Study:

• To evaluate the feasibility of herbicide susceptibility-based weed mapping using DCNNs.
• To facilitate targeted and efficient herbicide applications through advanced mapping.
• To optimize herbicide application paths using path-planning algorithms.

Main Methods:

• Implemented DCNNs (DenseNet, GoogLeNet, ResNet) for weed mapping based on herbicide susceptibility.
• Compared the performance of various DCNN models in terms of accuracy and efficiency.
• Applied path-planning algorithms (Christofides, Greedy, 2-opt) to optimize spraying nozzle trajectories.

Main Results:

• ResNet model demonstrated high accuracy (0.9980) and efficiency for weed detection.
• DenseNet achieved excellent F1 scores (0.992-0.999) across all herbicide categories.
• The Greedy algorithm was most efficient for optimizing nozzle path planning.

Conclusions:

• Herbicide susceptibility-based weed mapping enables precise herbicide application by targeting susceptible weeds.
• Integrating weed mapping with optimized path planning on smart sprayers can significantly reduce overall herbicide input.
• This approach enhances the efficiency and environmental sustainability of turf weed management.