<p>Ground penetrating radar (GPR) is the primary technology for detecting hidden cavities beneath urban roads. The number of effective cavities in actual detection scenarios is limited, leading to a severe shortage of training samples for intelligent underground defect recognition algorithms. Existing generative adversarial algorithms do not incorporate radar wave characteristics and lack directionality, making it difficult for the models to converge. To address this issue, this study proposes a directed generative adversarial network algorithm called GPR-CAE-GAN to expand the radar sample database of road cavities. This algorithm integrates easily accessible radar wave characteristics of underground targets with a small number of real measured cavity defect samples. An encoder is used to extract radar wave features that are easily detectable for underground targets, and these features are fused with noise and input into the generator. The discriminator is trained using real cavity data. Radar wave characteristics guide the algorithm in generating synthetic underground cavity radar data that resemble real scenarios, providing training data for underground target classification algorithms. The experiments validate the algorithm using real radar data in different geological environments, and the experimental results show that the samples generated by GPR-CAE-GAN are close to the real environments, which can improve the detection accuracy of underground target classification algorithms as training data. The algorithm provides a new idea to break through the limitation of insufficient training samples for the intelligent recognition algorithm of underground disease in GPR.</p>