The Global Wheat Full Semantic Organ Segmentation (GWFSS) dataset.

Area of Science:

• Agricultural Science
• Computer Vision
• Plant Phenotyping

Background:

• Computer vision and AI are vital for quantifying plant traits in agriculture.
• Current AI models struggle with segmenting complex wheat canopies.
• Accurate segmentation is crucial for monitoring crop health and development.

Purpose of the Study:

• To develop improved AI models for semantic segmentation of wheat organs (leaves, stems, spikes).
• To create a comprehensive dataset (Global Wheat Full Semantic Segmentation - GWFSS) for training and evaluating these models.
• To enhance the ability to distinguish wheat from weeds and identify senescent or diseased tissues.

Main Methods:

• Assembled a diverse global dataset (GWFSS) with 1096 pixel-level annotated images and 52,078 unannotated images.
• Trained segmentation models, including DeepLabV3Plus and Segformer, using the annotated dataset.
• Evaluated model performance based on mean Intersection over Union (mIOU) for different plant organs.

Main Results:

• The Segformer model achieved high mIOU (approx. 90%) for wheat leaves and spikes.
• Stem segmentation precision was lower (54%).
• The models successfully excluded weeds and identified necrotic/senescent tissues and crop residues.

Conclusions:

• The GWFSS dataset and trained AI models significantly advance semantic segmentation in wheat canopies.
• Improved segmentation facilitates accurate quantification of senescence and disease, addressing critical agricultural needs.
• The developed models offer superior performance in weed exclusion and tissue differentiation compared to existing methods.