Aiming at the problems that the existing decision tree methods rely heavily on prior knowledge in setting threshold values and the neural network methods in the field of communication signal recognition have large model sizes and high parameter counts, this paper proposes a lightweight neural network modulation recognition method integrating decision trees. This method introduces the idea of decision trees to analyze the confusion matrix of the dataset, divides the dataset into different subclasses based on the characteristics of different signal categories, and uses lightweight convolutional neural networks for hierarchical recognition. To achieve targeted recognition for each subclass in the hierarchical recognition process, data cleaning and feature extraction are employed to obtain the unique signal features of each subclass. Experimental results show that on the public dataset RML2016.10a, the overall recognition rate of this method reaches 90.03% within the signal-to-noise ratio range of 0 to +18 dB, which is 7.49% higher than the highest recognition rate of the comparison models. When the signal-to-noise ratio is 18 dB, the recognition rate reaches 95.03%; and the model parameter count is only 86 342, which is 96.85% lower than that of models with the same accuracy.