Weed mapping using UAV imagery and AI techniques: current trends and challenges.

Area of Science:

• Agricultural Engineering
• Computer Science
• Remote Sensing

Background:

• Unmanned aerial vehicles (UAVs) achieve high accuracy in weed recognition (>90%).
• Real-time weed identification on embedded systems remains a challenge.
• UAV-based weed mapping is crucial for precision agriculture.

Purpose of the Study:

• To review and analyze machine learning and deep learning (DL) applications for weed recognition using UAV imagery.
• To highlight methodologies, challenges, and advantages/limitations of current research.
• To guide future research in real-time weed mapping and site-specific herbicide application.

Main Methods:

• Systematic literature review of academic research on UAV-based weed recognition.
• Organization and comparison of studies based on methodology and addressed issues.
• Analysis of classical and DL approaches for feature extraction and classification.

Main Results:

• Classical methods focus on spectral, texture, and geometric features, with a trend towards non-visible spectra.
• DL methods excel at automatic multi-scale feature extraction directly from images.
• DL shows significant promise for distinguishing weed species and types.

Conclusions:

• UAV imagery combined with DL offers a powerful approach for weed mapping.
• Further research is needed to overcome challenges in real-time embedded systems.
• This review provides insights for developing intelligent systems for site-specific weed management.