Field boundary delineation with seasonal sentinel 2 imagery using Segment Anything Model (SAM)

Area of Science:

• Agricultural Science
• Remote Sensing
• Computer Vision

Background:

• Accurate field boundary delineation is crucial for precision agriculture (PA) and crop yield modeling.
• Traditional methods are labor-intensive and difficult to scale for large agricultural areas.
• High-resolution satellite imagery and advanced models offer potential for automated solutions.

Purpose of the Study:

• To develop and validate a fully automated workflow for large-scale field boundary extraction.
• To integrate the Segment Anything Model (SAM) with time-series Sentinel-2 imagery for enhanced segmentation.
• To provide a reproducible and adaptable methodology for PA applications.

Main Methods:

• Utilized a pre-trained foundation model, SAM, combined with Sentinel-2 time-series imagery.
• Generated seasonal composites of Red, Green, and Blue bands at various phenological stages.
• Developed a four-step workflow: environment setup, Google Earth Engine preprocessing, SAM segmentation, and ArcGIS Pro post-processing.

Main Results:

• Applied the workflow across over 32 million hectares in the Canadian Prairies.
• Achieved a high intersection-over-union (IoU) accuracy of 0.86 compared to manual segmentation.
• Demonstrated a scalable and efficient automated field boundary mapping solution.

Conclusions:

• The proposed automated workflow effectively delineates field boundaries at a large scale.
• This method provides a significant advancement for precision agriculture applications.
• The reproducible workflow is adaptable to different regions and datasets for automated mapping.