A Multi-Modal Dataset for Automated Phenological Stage Mapping in Actinidia chinensis.

Area of Science:

• Agricultural Science
• Computer Vision
• Plant Science

Background:

• Manual phenological monitoring of Actinidia chinensis is labor-intensive and limits precision agriculture scalability.
• Existing phenological datasets often lack spatial validation, hindering real-world application.
• Automated monitoring systems require comprehensive, spatially validated datasets for development and benchmarking.

Purpose of the Study:

• To introduce the Multi-Modal Actinidia chinensis Phenology Dataset, integrating annotated images and georeferenced videos.
• To provide a robust dataset for training computer vision models for phenological stage detection.
• To facilitate the spatial validation of automated counting algorithms and promote precision agriculture in kiwifruit production.

Main Methods:

• Developed a dataset comprising 1,665 annotated images and georeferenced videos of Actinidia chinensis.
• Utilized an adapted 17-class BBCH system, consolidating visually similar stages and introducing structural classes.
• Organized data hierarchically by plant structure, gender, and phenological stage, with manual ground truth for spatial distributions.

Main Results:

• The dataset enables the training of computer vision models for accurate phenological stage detection.
• Georeferenced videos allow for the validation of automated counting algorithms.
• The integrated approach achieves plant-level detection accuracy and offers a methodology for spatial validation.

Conclusions:

• The Multi-Modal Actinidia chinensis Phenology Dataset supports the development of automated phenological monitoring systems.
• This resource is crucial for optimizing operational costs and yield prediction in kiwifruit cultivation.
• The dataset promotes advancements in precision agriculture through enhanced spatial validation techniques.