A novel automated cloud-based image datasets for high throughput phenotyping in weed classification.

Area of Science:

• Agricultural Science
• Computer Vision
• Machine Learning

Background:

• Deep learning for weed detection requires extensive data management, including labor-intensive data acquisition and labeling.
• Limited temporal variation in existing datasets hinders the development of effective weed detection models.
• Robust weed identification models are crucial for precision agriculture and crop management.

Purpose of the Study:

• To introduce a cloud-based automatic data acquisition system (CADAS) for efficient image capture of crops and weeds.
• To address the challenges of data acquisition and temporal variation in datasets for weed detection.
• To provide a valuable, publicly released dataset for advancing deep learning in agriculture.

Main Methods:

• Developed a cloud-based automatic data acquisition system (CADAS) integrating fifteen digital cameras.
• Utilized gphoto2 libraries, external and cloud storage, and a Linux operating system for image capture.
• Captured images at fixed time intervals to account for plant growth stages and temporal variations.

Main Results:

• The CADAS system successfully captured images of six weed and eight crop species.
• A public dataset comprising 2000 images per species (raw and cropped with annotations) was released.
• The dataset facilitates research into deep learning-based weed and crop detection challenges.

Conclusions:

• The CADAS offers an automated solution to the time-consuming data acquisition phase in weed detection.
• The released dataset can help mitigate data imbalance issues and improve deep learning model performance.
• This work supports the advancement of intelligent agricultural systems and precision farming.