In order to optimize the fruit production, it is necessary to rationally remove some flowers during the flowering stage of apple trees. Orchards are usually thinned manually, which has problems such as being time-consuming and relying on personal experience. The purpose of this research is to apply a 3D reconstruction technique based on depth camera and semantic segmentation, and to obtain the position of flower clusters to be thinned based on semantic point cloud analysis to guide the robotic arm to thin the flowers. This research proposes real-time point cloud reconstruction algorithm and 3D spatial flower-branch relationship extraction algorithm and depends on the flower thinning strategy to decide whether to keep the flower clusters or not. The flower-branch relationship is calculated by clustering and skeletonising the flower and branch points in the semantic point cloud and locating the branch position of the flower cluster based on a 3D spatial search algorithm. The flower thinning strategy analyses the retention priority of each flower cluster based on the spatial relationships of the flower-branch. With this new method, precise flower thinning is achieved. In the future, we will calculate the position of the flower cluster based on the 3D model to guide the robot arm movement and execution.