The performance of ambient sound classification has substantially improved with the development of neural networks. The presence of noise in the classification of ambient sound remains a significant difficulty, nevertheless. In this research, we offer a method for classifying ambient noises under noise interference, with the main goal of increasing the classification accuracy of the model in the presence of noise interference, for the scenario where noise affects the classification effect. Prior to extracting the Mel frequency cepstrum parameter features and feeding them into the calibrated DenseNet121 model for classification training, the input ambient audio is first noise-reduced using the subspace approach. The more common five sound kinds were categorised using the experimental data to determine the environmental sounds gathered by the western black-crowned gibbon's acoustic monitoring system in the Mourning Mountains. The proposed method improves the accuracy of the training and test sets by 3.3% and 3.2%, respectively, and achieves a classification accuracy of 98.7%.