In view of the difficulty in hail monitoring and hail disaster estimation, combined with the characteristics of acoustic signal in time and frequency domain, this paper proposes a generalized regression neural network (GRNN) hail detection method based on entropy method by combining the entropy method with the GRNN. The time domain feature, frequency domain feature and wavelet packet energy spectrum feature are extracted from the collected hail and rain sound signals. The entropy method is used to determine the weight of each feature. The feature items with smaller weight are eliminated and fused to form a 〖JP2〗new feature subset. The feature subset is input into GRNN for prediction and recognition. The experimental results show that the method can effectively identify hail, and the recognition rate after feature screening is as high as 97.827 〖KG-*5〗6%, which is nearly 10% higher than that of the feature set without feature screening.