A DRCF-YOLO based method for rice seedling detection and density distribution mapping in UAV imagery

Accurate rice seedling detection and density distribution mapping from UAV imagery are essential for large-scale field monitoring and precision rice management. However, reliable detection remains challenging due to the small size of seedlings in high-resolution aerial images, subtle inter-class variations among seedling categories, and complex field backgrounds. Existing YOLO-based detectors often suffer from insufficient fine-grained feature extraction and limited multi-scale representation capability when applied to such small agricultural targets. To address these technical limitations, we propose DRCF-YOLO, a novel rice seedling detection model based on YOLOv11. A C3k2 module with Deformable Convolutional Network (C3k2-DCNv3) is incorporated into the backbone to enhance fine-grained feature extraction and discrimination across different rice seedling categories. Additionally, an enhanced Reparameterized Generalized Feature Pyramid Network (RepGFPN) combined with Convolutional Block Attention Module (CBAM) is introduced in the neck to strengthen multi-scale semantic fusion and improve small-target feature representation. Furthermore, the detection head is redesigned, incorporating Focal EIoU loss to alleviate sample imbalance. Experimental results show that DRCF-YOLO achieves an mAP@0.5:0.95 of 68.2% and a precision of 91.6% for Category II seedlings, exceeding the baseline model by 5.1% and 5.9%, respectively. Density distribution maps were then generated for three test plots through pixel–region coordinate transformation, yielding an RMAE of 4.1%. These results indicate that the proposed framework can accurately identify rice seedling status and convert discrete seedling locations into continuous spatial data, providing reliable support for precision agricultural practices such as variable-rate fertilization and targeted replanting.