Squall line (SL) is a convective weather process that often causes disasters. The automatic recognition and early warning of SL are important objectives in the field of meteorology. By collecting the new-generation weather RADARs (CINRAD/SA and CINRAD/SAD) base data during 12 SL weather events occurred in Jiangsu, Shanghai, Shandong, Hebei, and other regions of China from 2019 to 2021, the dataset has a total of 49,920 samples with a window size of 40 km. The 40 km area was labeled by employing manual classification and data augmentation to construct the deep learning dataset with a positive and negative sample ratio of 1:1, of which 80% and 20% are separated as the training and test set, respectively. Based on the echo height of each elevation beam at different distances, three deep learning-based models are trained for SL automatic recognition, which include a near-distance model (M1) trained by the data in nine RADAR elevation angles within 45 km from RADARs, a mid-distance model (M2) by the data in six elevations from 45 to 135 km, and a far-distance model (M3) by the data in three elevations from 135 to 230 km. A confusion matrix and its derived metrics including receiver operating characteristic (ROC) curves and the area under the ROC curve (AUC) are introduced as the indicators to evaluate the models by the test dataset. The results indicate that the accuracy of models are over 86% with the hit rates over 87%, the false alarm rates less than 21%, and the critical success indexes (CSI) surpass 78%. All the optimal critical points on the ROC curves are close to (0, 1), and the AUC values are above 0.95, so the three models have high hit rates and low false alarm rates for ensuring SL discrimination. Finally, the effectiveness of the models is further demonstrated through two SL events detected with Nanjing, Yancheng and Qingpu RADARs.