Compared to rainfall data observed at rain gauges, radar-based rainfall data can more effectively reflect the rainfall spatio-temporal distribution, which is of great significance for investigating the hydrological processes in a watershed and extending the lead time of flood forecasting. To explore the potential of radar in rainfall nowcasting, the radar echo image dataset from the Liulin experimental watershed were selected. The Farneback optical flow method and U-Net network model were employed to extrapolate radar echo images at different lead times. Quantitative precipitation estimation was conducted based on dynamic Z-R relation, and rainfall nowcasting results were compared and analyzed against the observations at rain gauges. The findings indicated that within a 30-minute lead time, the Farneback optical flow method exhibited superior performance, achieving a Probability of Detection (POD) of 0.933. Conversely, for lead times of 1 hour and 2 hours, the U-Net network demonstrated better nowcasting performance, with POD values of 0.956 and 0.948, respectively. The effectiveness of the Farneback optical flow method declined significantly with an extended lead time, while the performance of the U-Net network showed a less pronounced correlation with lead time.