Abstract: Due to the small scale and strong dispersion of lightning activity in plateau mountainous regions, it is difficult to implement accurate early warning of lightning disaster for key resource areas. Therefore, the correlation between the spatial and temporal evolution of thunderstorms and the characteristics of ground atmospheric electric field is taken into consideration, and a lightning warning methodology in mountainous regions based on ground atmospheric electric field apparatus and real-time lightning location system is proposed. By analyzing the evolution characteristics of atmospheric electric field under different thunderstorm development conditions in typical plateau mountainous regions, it is found that the atmospheric electric field can be used as a supplement to lightning location data, which fully represents the characteristics of severe discharge of thunderclouds and approaching development of thunderstorms. In the early warning process, the fast jitter and transient change characteristics from the morphological gradient of atmospheric electric field as well as the matched spatial and temporal distribution of cloud-to-ground lightning flash are input into the stacked sparse auto-encoder model, to indicate whether there is the existence of electrification in nearby thunderclouds. Secondly, the approaching trend of thunderstorm activity is determined by the thunderstorm distance variation or the electric field waveform pattern. Finally, the upcoming lightning activity in a localized area with a radius of 15 km is forecast based on the combined results. The case study in the mountainous regions of Yunnan province during the thunderstorm season of 2023 reveals that the effective identification of the early warning features of lightning activity extracted from the dual-source data can achieve the early warning accuracy of 90%, and about 44% of the warning time leads at least 30 min.