Abstract:The serious imbalance in sample distribution, characterized by a sharp drop in the frequency of severe convective winds with increasing wind speeds, is identified as the predominant factor hindering the accurate intensity-based classification of severe convective winds by various existing algorithms.To address this problem, in this study the non-differentiable probability of detection (POD) is proposed to be the loss function for a convolutional neural network (CNN) and Bias to be its constraint condition. Subsequently, the multi-objective optimization immune evolution algorithm (MOIEA) is employed to optimize all the model parameters of the CNN. This contributes to the development of a novel identification algorithm, which is named severe convective wind identification network (SCWINet), for identifying severe convective winds at the speeds of 17.2, 20.8, 24.5 m·s-1 and above. SCWINet leverages the radar vertical liquid water content, three-dimensional radar reflectivity, lightning location data and minutely surface automatic observation station data in Zhejiang Province during 2022-2024, achieving different levels of severe convective wind identification with temporal resolution of 6 minutes and spatial resolution of 0.01°. Then, the performance of SCWINet is compared to weighted mean-square error (WMSE) and balanced mean-square error (BMSE), which use the same CNN structure but have differentiable loss functions. The applicability of SCWINet is then assessed based on the threat score, Bias, POD, false alarm ratio that uses the neighborhood method (with a scanning radius of 5 km), and the planar distribution characteristics of severe convective winds. The main results are as follows. SCWINet can effectively identify severe convective winds of 17.2, 20.8, 24.5 m·s-1 and above corresponding to systematic and scattered severe convective systems, with better performance observed in identifying severe convective winds triggered by systematic convection than those triggered by scattered convection. However, the identification effectiveness of SCWINet generally decreases as wind speed increases, with increased false alarms and missed detections being the primary causes of this phenomenon. By contrast, the commonly used WMSE and BMSE approach-es fail to identify severe convective winds, and all severe convective winds they identify are below 17.2 m·s-1.Nevertheless, the data used in this study are somewhat limited in terms of the feature completeness and volume. Future enhancements in identification accuracy of severe convective winds could be achieved by incorporating additional features and data, such as radar radial velocity, specific differential phase, differential reflectivity, and satellite data. This could be also applied to identify even higher wind speeds.