The objective of this study is to effectively improve the timeliness and accuracy of severe convective weather early warning in the mountainous areas of central Yunnan under complex terrain conditions. Firstly, the quality of multi_type radar data is improved based on the dual_polarization radar parameters. Secondly, a collaborative networking method for multiple dual_polarization weather radars is established to obtain fusion data with consistent spatio_temporal resolution. Finally, a fast and accurate early warning method for severe convective weather is built based on macro_ and micro_scale radar characteristics data of classified severe convective weather, surface station observation data, and an improved ViT_Large deep learning model. In addition, comparative analysis experiments and quantitative evaluations are conducted with historical data of severe convective weather processes in the central Yunnan region over the past five years. The results show that, relative to traditional early warning methods, the improved intelligent early warning method helps achieve the following improvements: the critical success index mean score is increased by approximately 58. 5%; the false alarm rate mean score is decreased by approximately 29. 75%; the missing alarm rate mean score is decreased by approximately 40%; the continuous ranked probability mean score is decreased by approximately 51. 42%; and the lead time of early warnings is increased from an average of approximately 9. 5 min to approximately 27 min. Therefore, the intelligent early_warning method for severe convective weather established in this study can cover the blind areas of observation caused by mountain terrain, improve the spatio_temporal resolution of observation data, and solve the problem of non_linear development and evolution of severe convective weather under complex terrain conditions. This could significantly enhance the timeliness and accuracy of early warning of severe convective weather in the mountainous regions of central Yunnan.