SPVD-field: a task-oriented multi-task visual dataset for sweet potato virus disease under real field conditions

Area of Science:

• Agricultural science
• Computer vision
• Plant pathology

Background:

• Sweet potato virus disease (SPVD) causes significant global yield losses, threatening food security.
• Vision-based intelligent diagnosis offers a scalable solution for SPVD monitoring.
• Existing datasets are limited, hindering advanced computer vision method development for SPVD analysis.

Purpose of the Study:

• Introduce SPVD-Field, a novel task-oriented multi-task visual dataset suite for SPVD.
• Facilitate the development, evaluation, and comparison of computer vision methods for SPVD.
• Support diverse research in detection, segmentation, and phenotyping.

Main Methods:

• Developed SPVD-Field, comprising two complementary sub-datasets: SPVD-DET (detection) and SPVD-SEG (segmentation).
• Collected data independently with optimized protocols for each task, ensuring a unified semantic definition.
• Captured real-world variability in imaging conditions, including scale, viewpoint, illumination, and background complexity.

Main Results:

• SPVD-Field provides bounding-box annotations for detection and pixel-level masks for segmentation.
• Demonstrated dataset usability and difficulty through baseline benchmark results for detection and segmentation.
• Documented data acquisition, annotation, and quality control procedures.

Conclusions:

• SPVD-Field addresses the critical need for comprehensive, task-oriented datasets in SPVD research.
• The dataset enables reproducible and comparable research in plant phenotyping and disease analysis.
• Facilitates advancements in multi-task learning and disease severity assessment for sweet potatoes.