Multi-stage progressive detection method for water deficit detection in vertical greenery plants.

Area of Science:

• Agricultural Science
• Computer Vision
• Plant Physiology

Background:

• Accurate detection of water deficit in vertical greenery plants is crucial for cultivation.
• Current single-stage target detection methods struggle with complex environments and image quality variations.
• Existing methods lack efficiency and accuracy in real-world vertical greenery plant monitoring.

Purpose of the Study:

• To develop a multi-stage progressive detection method for enhanced accuracy and efficiency in identifying water deficit in vertical greenery plants.
• To introduce a Swin Transformer with mobile windows and hierarchical feature extraction for improved computational performance.
• To address the limitations of current detection algorithms in complex cultivation scenarios.

Main Methods:

• Proposed a multi-stage progressive detection architecture for gradual image filtering, processing, and detection.
• Integrated a Swin Transformer with mobile windows and hierarchical representations for efficient feature extraction.
• Employed a self-attention mechanism for global feature modeling to enhance detection capabilities.

Main Results:

• The multi-stage detection approach achieved a high average precision of 93.5% for vertical greenery plants detection.
• Demonstrated significant accuracy improvements over Mask R-CNN (19.2%), YOLOv7 (17.3%), DETR (13.8%), and Deformable DETR (9.2%).
• The method effectively reduced computational load while enhancing overall detection efficiency.

Conclusions:

• The proposed multi-stage progressive detection method offers a superior solution for accurate and efficient water deficit detection in vertical greenery plants.
• The integration of Swin Transformer and self-attention mechanisms contributes to robust feature extraction and global modeling.
• This advancement has significant implications for precision agriculture and sustainable urban greening practices.