Abstract: It is an important method for classifying UAVs detected by radar using convolutional neural networks to perform deep learning on targets’ micro-Doppler feature. Actually, parameters of UAVs such as blade rotation speed, blade length, blade initial phase, UAVs’ azimuth, UAVs’ pitch angle, and UAVs’ radial velocity, etc. vary greatly, which leads to large variation in training samples sets. In this paper, impacts of training samples sets on rotor UAVs’ classification results are analyzed. Firstly, simulated radar echoes models of helicopters, quadrotors and hexarotors are established. Then micro-Doppler features analysis and extraction are carried out, and Merged Doppler Images (MDI) training samples sets are constructed in many different situations. Finally, GoogLeNet (Inception v1) is used to obtain UAVs’ classification results in different situations. Impacts of sample quantity, variation of UAVs’ single parameter, completeness of sample parameters coverage and sampling intervals of UAVs’ parameters of training sets on the classification accuracy are analyzed. The experiment results show that the difference in MDI training sets may have significant impacts on UAVs’ classification accuracy.