Urban tree detection is crucial for effective city tree management. Recently, deep learning methods have become mainstream for urban tree detection. However, these methods lack sufficient training samples for the diverse types of urban trees and struggle to handle occluded trees. To address these problems, we propose a simple yet effective Copy-Paste data augmentation method for urban tree detection. To address the lack of diverse urban tree training samples, we create an extra urban tree image library and use our Copy-Paste strategy based on this image library. To avoid unrealistic random pasting, we use ViT-Adapter to segment images to obtain the category for each region in the image. In each iteration, we randomly select a point within three categories of regions and paste an image randomly chosen from our image library. To address the problem of poor model performance in complex occlusion scenarios, we introduce more overlapping trees, with an overlap threshold between each pair of pasted trees. Experimenting on the UTD dataset and applying our method to four representative object detection algorithms, we observe a significant improvement in performance. Our method does not increase the parameters and Giga Floating-point Operations Per Second (GFlops) of the original models, and it has almost no impact on frames per second (FPS).