A deep learning approach based on graphs to detect plantation lines.

Area of Science:

• Computer Vision
• Agricultural Technology
• Machine Learning

Background:

• Automatic farming requires accurate identification of plantation lines in aerial imagery.
• Existing deep learning methods struggle with complex plantation patterns, especially those with spaced plants.
• UAV-based RGB imagery presents unique challenges for plantation line detection.

Purpose of the Study:

• To develop a robust deep learning approach for detecting plantation lines in challenging aerial imagery.
• To address limitations of current methods in handling spaced plant patterns and plantation gaps.
• To improve the automation of farming processes through precise line extraction.

Main Methods:

• A graph-based deep learning model utilizing VGG16 backbone for feature extraction.
• A Knowledge Estimation Module (KEM) to detect plant positions, plantation lines, and displacement vectors.
• Graph modeling where plant positions are vertices and edges are classified based on visual features, KEM probabilities, and vector alignment.

Main Results:

• The proposed method demonstrated superior performance compared to state-of-the-art deep learning techniques across corn, orange, and eucalyptus datasets.
• Individual module advantages were validated through experiments on corn plantations.
• The approach effectively extracts lines in spaced plantation patterns, handling interruptions and gaps.

Conclusions:

• The graph-based deep learning method offers a significant advancement in detecting plantation lines, particularly in scenarios with spaced plants.
• This technique enhances the reliability of automatic farming processes by providing accurate line extraction.
• The method shows strong generalization capabilities across different crop types and growth stages.