A Multi-Head UNet++ Framework with Fractional Differential Output Refinement for UAV Multispectral Crop Stress...

Area of Science:

• Agricultural Science
• Computer Vision
• Remote Sensing

Background:

• Traditional crop stress mapping often analyzes water stress and rust disease separately.
• Integrating diverse datasets with varied annotation schemes presents a significant challenge.
• Developing unified frameworks is crucial for comprehensive agricultural monitoring.

Purpose of the Study:

• To present a unified semantic segmentation framework for UAV-based multispectral crop stress mapping.
• To integrate water stress and rust disease detection into a common label space.
• To harmonize heterogeneous datasets for a single multi-class segmentation problem.

Main Methods:

• Utilized a patch-based strategy for processing UAV multispectral orthomosaics.
• Employed a multi-head UNet++ architecture with segmentation, edge-aware, and Signed Distance Transform (SDT) branches.
• Introduced a physics-informed output-space refinement module based on fractional partial differential equations (FPDE).

Main Results:

• Demonstrated effective boundary delineation and spatial consistency in predicted stress maps.
• Showcased improved detection of minority stress classes.
• Validated the framework's effectiveness within the evaluated dataset setting.

Conclusions:

• The proposed framework successfully integrates heterogeneous crop stress conditions into a unified segmentation approach.
• The method enhances spatial coherence and boundary preservation in stress mapping.
• This research lays the groundwork for scalable, multi-source agricultural monitoring systems.