Abstract: Based on data of real-time precipitation and weather radar during 28 torrential rain events in warm regions of Sichuan Basin, radar echo characteristics of torrential rain in early and mature stages are analyzed. Feature vectors for identifying early and mature stages of torrential rain in warm regions are constructed and the selected samples are studied by random forest machine learning method. According to the influence range, duration, and cumulative amount of precipitation, the thunderstorm group is the main part of the rainstorm in the warm region, and its development can be divided into three types.According to the burst of short-term heavy precipitation, thunderstorm groups are divided into primary and mature stages. In the early stage of the thunderstorm to the mature stage, the "in situ development type" is dominant, the "individual development type" and the "in front side trigger type" are the second. With the evolution after the maturity, the "in situ development type" and the "front side trigger type" are the main types. Convective precipitation is the main type of heavy rain in the warm area. After the first type of thunderstorm group, the combination of mature thunderstorms is the main source of thunderstorms, which moves slowly and is conducive to the generation of heavy precipitation. In front of mature thunderstorms, new thunderstorm cells are continuously generated and merged to continue spreading northward, forming a large range of precipitation. Individual development thunderstorm groups have the longest duration and a large influencing range. They are accompanied by long-term merge when moving. Among 28 processes, a large proportion appears in the northwest and has the longest duration. Echoes of these processes are in the southwest-northeast direction, which is basically consistent with the trend of the Longmen mountains in the western part of the Basin. The uplift of the topography (generating easterly wind) plays a key role in the occurrence and development of warm rainstorms. In the primary stage, the average core height and average top height of "in situ development type" thunderstorm group has a bimodal structure. Similar structure is found for "front side trigger type" in the mature stage. Multiple parameters of three types of thunderstorm groups show a unimodal distribution in the nascent and mature stages. To identify heavy rains in the warm area, feature vector is constructed using multiple parameters of the thunderstorm group, and random forest machine learning is also applied, leading to satisfying results.