Deep learning–based object detection for automated classification of ripe and unripe tomatoes in field conditions

Accurate detection of tomatoes at different ripening stages from close-range remote sensing imagery acquired under real field conditions is a key enabler for yield estimation, harvest scheduling, and selective robotic picking in precision agriculture. However, achieving reliable performance is challenging due to occlusions, variable illumination, dense foliage, and the limited size of domain-specific datasets typically available to small farms and research teams. In this work, we present a controlled comparative study between two representative object detection paradigms, Faster R-CNN with a ResNet-50 FPN backbone (two-stage) and YOLO26s (single-stage), on a curated dataset of 367 high-resolution images acquired in a commercial cherry-tomato greenhouse under natural illumination, containing more than 8,900 manually annotated tomatoes across three classes: blossom, premature, and ripe. Both models were trained for 100 epochs under a controlled protocol sharing the same data, stratified 70/15/15 split complemented by 5-fold cross-validation, and evaluated using COCO-style AP/AR together with per-class precision–recall and ROC analyses. YOLO26s achieved a macro-average AP of 0.741 and ROC AUC of 0.954 on the held-out test split, substantially outperforming Faster R-CNN (macro AP 0.241; ROC AUC 0.836), particularly under stricter localization thresholds, and qualitative examples confirm that localization rather than classification is the main bottleneck of Faster R-CNN. These results provide actionable guidelines for deploying detectors on small agricultural remote sensing datasets in fixed monitoring stations and on ground or aerial autonomous platforms, where tighter bounding boxes and a favorable accuracy–efficiency trade-off are decisive for reliable operation.