LoRa technology plays a vital role in IoT applications by enabling low-power and long-range communication. Since the LoRa signal may transmit hundreds of meters to tens of kilometers under different environments, estimating the path loss of LoRa links can bring great benefits to the network, such as device localization and network planning. Recent years have seen some mechanisms to improve the LoRa link estimation by analyzing the environments based on satellite images. However, they may be inaccurate when we consider the actual network in the three-dimensional (3D) scene. This paper proposes 3DLoRa, the first framework to refine LoRa link estimation through exploiting 3D point clouds. We generate a 3D map by conducting semantic segmentation and voxelization based on the point clouds. We then propose both model-based and learning-based methods to estimate the path loss of a link according to the environments obtained from the 3D map. We test the 3D map generation process based on 3D point clouds in a campus area and show an accuracy of about 97% for environment categorization. We also deploy a LoRa network on the campus and collect data to evaluate the performance of link estimation. Experimental results indicate that 3DLoRa can significantly improve the accuracy of link estimation in various environments compared to previous works. Specifically, the model-based method can reduce the estimation error to about$7dBm$with low computational overhead, while the learning-based method can reduce the estimation error to about$2.5dBm$with the cost of higher computational overhead. These results can serve as a reference for how to deploy link estimation methods in different network situations.