DeepWeeds: A Multiclass Weed Species Image Dataset for Deep Learning.

Area of Science:

• Agricultural Robotics
• Computer Vision
• Machine Learning

Background:

• Robotic weed control offers potential for increased agricultural productivity but faces challenges in rangeland environments.
• Current research primarily focuses on croplands, neglecting the specific needs of rangeland stock farmers.
• Robust weed species classification in natural environments is a significant barrier to robotic weed control adoption.

Purpose of the Study:

• To introduce the first large, public, multiclass image dataset of weed species from Australian rangelands.
• To enable the development of robust weed classification methods for robotic weed control.
• To establish a baseline for classification performance using deep learning models.

Main Methods:

• The creation of the DeepWeeds dataset, comprising 17,509 labeled images of eight significant weed species across eight locations in northern Australia.
• Utilizing benchmark deep learning models, Inception-v3 and ResNet-50, for weed species classification.
• Evaluating the real-time performance of the ResNet-50 architecture for inference speed.

Main Results:

• Deep learning models, Inception-v3 and ResNet-50, achieved high average classification accuracies of 95.1% and 95.7%, respectively.
• The ResNet-50 model demonstrated real-time performance with an average inference time of 53.4 ms per image.
• The developed dataset and baseline results show promise for future field implementation of robotic weed control.

Conclusions:

• The DeepWeeds dataset is a valuable resource for advancing robotic weed control in Australian rangelands.
• Deep learning models show significant potential for accurate and efficient weed species classification in complex environments.
• The findings support the viability of robotic weed control systems for rangeland stock farmers.