Abstract: To conduct a more refined assessment of rainstorm disasters at prefectural-level city scale in Jiangxi Province, this study utilizes precipitation data from 1961 to 2020, rainstorm disaster data from 1984 to 2020, and GDP data to investigate the characteristics of rainstorm events and their disaster loss assessment models for 11 prefectural-level cities in Jiangxi. Firstly, it systematically analyzes the temporal and spatial distribution characteristics of rainstorm events and their associated disasters across the 11 prefectural-level cities. Secondly, regionally differentiated factors are screened using the significant correlation coefficient method to construct a comprehensive rainstorm intensity index (I) and disaster loss assessment models for 11 prefectural-level cities. Finally, the exponential regression method and the extreme gradient boosting method (XGBoost) are employed to build disaster loss grade assessment models, with model performance validated using training samples and independent samples. The results are as follows. (1) Obvious differences exist in the characteristics of rainstorm events and disasters among 11 prefectural-level cities in Jiangxi.Shangrao, Jingdezhen, and Yingtan are areas where rainstorm events occur frequently and severely. The frequency, cumulative rainfall and disaster frequency in Jiangxi have a significant upward trend with the increase of years, but the disaster loss rate shows a weak downward trend. (2) The number and types of significant correlation factors in different prefectural-level cities have obvious differences, among which the maximum process rainfall of a station, cumulative rainfall and duration are the most common three significant correlation factors. The I is significantly correlated to the loss rate in 11 prefectural-level cities. And the loss assessment models exhibit good fitting results for most prefectural-level cities, particularly for Fuzhou and Nanchang, with determination coefficients (R²) of 0.82 and 0.80, respectively. (3) For the disaster loss grade assessment models, testing on training samples shows that errors of both methods are within 2 levels, with XGBoost performing better in most prefectural-level cities.Testing on independent samples indicates that the model errors were basically within the one level, and the exponential regression method demonstrates better stability. Moreover, combining these two methods could accurately provide the level range intervals of rainstorm disaster losses.