Tomato Multi-Angle Multi-Pose Dataset for Fine-Grained Phenotyping.

Area of Science:

• Plant Science
• Computer Vision
• Agricultural Technology

Background:

• Traditional plant phenotyping methods suffer from observer bias and inconsistencies, limiting accuracy and reproducibility.
• Fine-grained plant analysis requires precise and reliable data acquisition and processing.

Purpose of the Study:

• Introduce TomatoMAP, a comprehensive dataset for Solanum lycopersicum (tomato) to overcome limitations in plant phenotyping.
• Develop and benchmark AI models for accurate and efficient automated fine-grained plant analysis.

Main Methods:

• Collected 68,080 RGB images of tomato plants, including high-resolution macrophotographs and moderate-resolution images from various poses and elevations.
• Annotated images with bounding boxes for seven regions of interest and 50 BBCH phenological growth stages.
• Prioritized and benchmarked real-time applicable AI models (MobileNetv3, YOLOv11, Mask R-CNN) using accuracy, mAP, and inference FPS.

Main Results:

• AI models trained on TomatoMAP demonstrated accuracy comparable to human domain experts.
• Benchmarking showed trade-offs between accuracy and efficiency for different AI models.
• Reliability of automated phenotyping was validated using Cohen's Kappa statistics and inter-rater agreement heatmaps.

Conclusions:

• TomatoMAP dataset significantly advances automated fine-grained plant phenotyping for Solanum lycopersicum.
• AI models trained on TomatoMAP offer a reliable and reproducible alternative to traditional methods.
• The dataset and models support real-time applications in precision agriculture and plant research.