Abstract Observer bias and inconsistencies in traditional plant phenotyping methods limit the accuracy and reproducibility of fine-grained plant analysis. To address these limitations, TomatoMAP is introduced as a comprehensive dataset for Solanum lycopersicum . The dataset contains 68,080 RGB images: 3,616 high-resolution macrophotographs (3648 × 5472) with semantic annotations, and 64,464 moderate-resolution images (1080 × 1440) captured from 12 plant poses at four camera elevations. Each image is accompanied by manually annotated bounding boxes for seven regions of interest (leaves, panicle, flower clusters, fruit clusters, axillary shoot, shoot, and whole-plant area) and by labels spanning 50 BBCH classes representing phenologically growth stages. A general cascading structure is proposed. For real-time applicability, models emphasizing the accuracy-efficiency trade-off (MobileNetv3, YOLOv11, and Mask R-CNN) are prioritized and benchmarked against multiple state-of-the-art models. Performance is assessed using accuracy, mAP, inference FPS, and normalized confusion matrices. In a study involving five domain experts, AI models trained on TomatoMAP achieves comparable accuracy levels. Reliability of automated fine-grained phenotyping is supported by Cohen’s Kappa statistics and inter-rater agreement heatmaps.
Zhang et al. (Sat,) studied this question.