Automated open-field plot grid extraction using zero-shot learning

Overview

The proposed solution utilizes a zero-shot learning approach to automate the extraction of open-field plot grids, particularly in vegetable trials. This innovative method leverages AI-driven image recognition to replace traditional manual plot delineation, significantly enhancing efficiency, accuracy, and reliability. By employing the Segment Anything Model (SAM) for semantic segmentation, this approach ensures consistent plot delineation at the field scale, even when dealing with poorly aligned orthomosaic images from UAVs lacking RTK GPS. The integration of this technology reduces labor and time, addresses irregular planting patterns, and improves the overall integrity of data collected during agricultural trials.

Technical specifications

Key features:

• Zero-shot learning capability allows semantic segmentation without the need for training data.
• Compatible with UAV-generated orthomosaic images, even when RTK GPS is unavailable.
• The framework includes three main tasks: mask generation, plot orientation estimation, and plot refinement.
• Utilizes the Segment Anything Model (SAM) for stable performance across diverse agricultural patterns.
• Potential enhancements include integrating multispectral bands and canopy height models.

Technology readiness level

Currently at Technology Readiness Level 5, this solution has been validated through UAV data collection and is in the development phase for automated plot grid extraction. Future plans include applying this framework to various crop types and enhancing capabilities to handle diverse agricultural scenarios.