Accurately detecting the type and distribution of underground targets is a prerequisite and key to ensuring the safety of urban underground space. Ground Penetrating Radar (GPR) is a conventional method for identifying concealed diseases in urban underground spaces. However, its data interpretation primarily relies on the manual work of experienced practitioners, which is time-consuming and labor-intensive, and deep learning-based automatic target recognition methods require a large number of labeled samples, mostly utilizing B-scan data, resulting in low recognition accuracy. To address these challenges, this paper proposes a three-dimensional (3D) GPR target recognition method based on temporal ensembling semi-supervised deep learning. This method combines three slices of 3D GPR data (B-scan, C-scan, and D-scan) into triple-channel GPR images as network inputs. By employing a temporal ensembling semi-supervised deep learning approach, the model is jointly trained using a small amount of labeled data and a large amount of unlabeled data, thereby reducing the demand for labeled data. The temporal ensembling method is a kind of self-ensembling learning method, which utilizes multiple versions (ensembles) of a model during training to enhance performance, robustness, and generalization without requiring much labeled data. Additionally, a triplet attention module is introduced to learn spatial and channel information of 3D data more effectively, thus improving the classification performance. Indoor experimental results demonstrate that when only 10% of the training data is labeled, the proposed method achieves an average accuracy of 96.45% for four types of underground targets (i.e., metal pipe, plastic pipe, void, and background). Comparative experiments indicate that this method improves the average accuracy by 5.28% compared to B-scan-based methods and outperforms supervised methods based on transfer learning and four semi-supervised methods. These findings demonstrate that the proposed method can be effectively applied to intelligent GPR underground target recognition with limited labeled data, facilitating the promotion of deep learning in GPR applications.