Feature extraction is one of the key technologies for aerial target attitude estimation. Traditional methods rely on manual feature extraction, which only uses shallow structural features, thus limiting the accuracy of target attitude estimation. In this paper, for the first time, a deep learning method is adopted to conduct research on aerial target attitude estimation based on radar cross section (RCS). Considering the small dimension of RCS data and the complex mapping relationship between RCS and attitude, one-dimensional Cut-ResNet50 with 6 Bottleneck is proposed by pruning ResNet50, which simplifies the model while increasing the generalization performance of the model. The loss function is modified by combining cross-entropy and mean square error to enhance the learning ability of the model. Finally, attitude recognition testing results on simulated RCS dataset and comparison with various models validate the effectiveness of the proposed method.