Abstract: Abstract: Densely planted seedlings in the nursery with overlapping canopies and large differences in size, have made counting seedlings difficult, if performed manually. Inaccurate seedling counting usually causes a nursery manager to make decisions mismatching with the existing state, thereby resulting the losses. It is necessary to develop automatic techniques for seedling counting, further to avoid the loss that caused by inaccurate seedling counting. In this study, the spruce plant images were collected by the unmanned aerial vehicles (DJI Phantom 4), while, a spruce image dataset was constructed. 558 spruce plant images with diversity were selected, and 20 complete plot images were stitched using Pix4D mapper software. In images, the contrast, angle, and size were adjusted to expand spruce images to 4 times of original images. The training set and test set were built according to the ratio of 7:3. An improved YOLOv3 model can quickly and accurately detect the targets with large size differences, such as spruce. However, in a small sample of spruce plants, the training process was prone to overfitting, where only a few dimensional features were used in the feature extraction process, resulting in the loss of spruce feature information and less counting accuracy. The YOLOv3 model was also verified. 1) Densely connected module was added to feature extraction network of YOLOv3 model, and the transfer and reuse of spruce features were strengthened. The number of model parameters was reduced to suppress the overfitting problems; 2) Transition module was added to feature extraction network of improved YOLOv3 model. The spruce feature information was extracted and fused using filters with different sizes and pooling operations to avoid spruce feature loss. Five evaluation indicators including precision, recall, average precision, mean counting accuracy, and average detection time were used to evaluate the counting. Five evaluation indicators in the improved YOLOv3 model were 96.81%, 93.53%, 94.26%, 98.49%, and 0.351 s, respectively. The improved YOLOv3 model can quickly and accurately realize spruce counting. Compared with the original YOLOv3 model, SSD model, and Faster R-CNN model, the improved YOLOv3 has significant advantages in 5 evaluation indicators. In the spruce images of a complete plot after stitching, five evaluation indicators in the improved YOLOv3 model were 91.48%、89.46%、89.27%、93.38%, and 1.847 s, respectively. Compared with the original YOLOv3 model, SSD model, and Faster R-CNN model, the performance of new model has significantly improved. The results demonstrated that the counting result of improved YOLOv3 model was greatly optimized, and further to make a useful exploration for UAV.