Maize Seedling Detection Dataset (MSDD): A Curated High-Resolution RGB Dataset for Seedling Maize Detection and Benchmarking with YOLOv9, YOLO11, YOLOv12 and Faster-RCNN

Seed germination and early survival are important phenotypes for plant breeding and agricultural management, yet they are still commonly assessed through labor-intensive manual stand counting. We present the Maize Seedling Detection Dataset (MSDD), a curated high-resolution red–green–blue (RGB) dataset derived from unmanned aerial vehicle (UAV) imagery collected over the 2019–2022 growing seasons. MSDD contains 3152 images and 163,921 annotated objects across three classes—single (92.47%), double (6.07%), and triple (1.45%) clusters of seedlings—and captures substantial variability in growth stage (V2–V12), illumination, soil appearance, wind, and camera viewpoint. Unlike many existing datasets, MSDD explicitly annotates clustered seedlings as double and triple classes, which are important for stand evaluation. We benchmarked YOLOv9, YOLO11, YOLOv12, and Faster-RCNN on MSDD to evaluate detection accuracy, class-specific performance, inference efficiency, and generalization across field conditions. Single-seedling detection was reliable across models, with the best mean average precision at 0.5 IoU (mAP@0.5) reaching 0.916, whereas double- and triple-seedling detection remained challenging because of class imbalance, occlusion, and annotation ambiguity. Detection was most reliable in high-contrast scenes and declined under wind, strong shadows, and bright soil backgrounds. YOLO11 provided the fastest evaluation throughput among the tested models (≈27 frames per second (fps)), while YOLOv9 achieved the strongest single-seedling detection performance. Synthetic augmentation improved class balance but did not improve generalization to naturally occurring clustered seedlings. Frames, labels, and trained models are available at Google Drive and Hugging Face. MSDD provides a public benchmark for maize seedling detection and for evaluating stand counting models under realistic field conditions.