Unmanned Aerial System-Based Weed Mapping in Sod Production Using a Convolutional Neural Network

Area of Science:

• Agricultural Science
• Computer Science
• Remote Sensing

Background:

• Weed control is a major challenge in sod farming, relying heavily on chemical herbicides.
• Advancements in unmanned aerial systems (UAS) and artificial intelligence offer new solutions for weed management.

Purpose of the Study:

• To investigate weed composition and area on sod farms using ground and UAS-based surveys.
• To train a convolutional neural network (CNN) for identifying and mapping weeds in sod fields using UAS imagery.
• To evaluate the performance of the CNN against human visual identification.

Main Methods:

• Conducted ground and UAS-based weed surveys to assess weed type composition and area.
• Utilized UAS imagery and the Fastai library (PyTorch) to train a CNN model.
• Compared CNN performance metrics (precision and recall) with human visual assessment.

Main Results:

• The CNN demonstrated comparable, and in some cases superior, performance to human identification, particularly for broadleaf and spurge weeds.
• Achieved high precision (0.68-0.88) and recall (0.78-0.99) across various weed classes (broadleaf, grass, spurge, sedge, no weeds) at specific resolutions.
• Showcased CNN's potential for high-accuracy weed detection (>0.9 precision/recall) during turf establishment with mature weeds.

Conclusions:

• CNNs trained with UAS imagery are effective tools for weed mapping in sod farms.
• The accuracy of weed detection is influenced by image resolution, suggesting potential need for multiple models.
• This technology can aid in optimizing herbicide application and improving weed management strategies.