Abstract: Boundary layer convergence lines are recognized as one of the critical mesoscale weather systems triggered convection, and also affect low-altitude flight safety. The accurate and detailed identification of these lines is considered essential for revealing their formation, evolution, and interaction mechanisms with other weather systems. However, existing automatic identification technologies are limited in their ability to adapt the diverse characteristics of these lines, such as scale, intensity, and shape. The rotational region-based convolutional neural network (R2CNN) is employed to enhance the accuracy, robustness, and generalization of the identification process. A comprehensive identification dataset has been constructed for model training, considering the diversity of weather radar models and resolutions. Relevant parameters are adjusted to derive the optimized recognition model. The intersection over union (IoU) with confidence levels are employed to comprehensively assess and validate the identification results. Results indicate that the boundary layer convergence line recognition algorithm developed achieves a higher hit rate and a lower false alarm rate at lower IoU thresholds. At a confidence level of 0.7, the threat score (TS) reaches its maximum value.Compared to the existing Machine Intelligence Gust Front Algorithm (MIGFA), the model proposed in this study demonstrates significant advantages in reducing false alarms, improving hit rates, and achieving a balanced recognition frequency. Therefore, it is more suitable for operational applications and dissemination. This research not only provides a more effective method for identifying boundary layer convergence lines but also contributes to the improvement of low-altitude flight safety and advances meteorological detection technologies. The proposed method addresses limitations of existing technologies by effectively managing the diverse characteristics of boundary layer convergence lines. By incorporating rotational bounding boxes in the detection process, R2CNN model enhances the detection accuracy for objects with arbitrary orientations, which is particularly beneficial for meteorological phenomena that do not align with the standard axis. The constructed dataset includes a diverse collection of radar images from various models and resolutions, ensuring that the model is trained on a wide range of data and can generalize effectively to new, unseen data. Extensive experiments are conducted to evaluate the model's performance under different IoU thresholds and confidence levels. Findings demonstrate that at lower IoU thresholds, the model maintains high detection performance, indicating its robustness in practical applications where precise localization may be challenging. Furthermore, the superior performance of the proposed model compared to MIGFA indicates its potential for widespread adoption by meteorological agencies for better monitoring and forecasting.