Aiming at the ground-based synthetic aperture radar (SAR) deformation measurement, the conventional permanent scatterer (PS) selection method is difficult to meet the deformation measurement requirements in terms of the quantity and quality of PS selection in the time-incoherent and complex scenarios. The article proposes a PS selection method based on bi-directional long short-term memory-convolutional neural network (BiLSTM-CNN), which uses amplitude dispersion and amplitude to jointly select positive and negative samples to construct the training dataset, and takes the interfering phase, amplitude divergence and correlation coefficient as the temporal features of the dataset, and then learns the PS global temporal features and the PS local temporal features by using the BiLSTM and the multi-scale CNN, respectively,and then the global and local temporal features are weighted and fused to learn by the Multi-Head Self-Attention (MHSA), and finally feature probability mapping is carried out in order to construct the PS classification model.The performance of the proposed selection method is experimentally analyzed by using the radar monitoring data of Jiudaoquan in Wanzhou District, Chongqing Municipality, and the results show that the method improves the network accuracy, score, recall, precision, and other indexes, and improves the quantity and quality of radar image PS selection.