In recent years，the widespread use of drones in various fields has introduced new security risks，making the monitoring and detection of drone activities crucial. This paper proposes a UAV acoustic detection and recognition method based on the mel spectrogram and time-frequency attention and soft thresholding convolutional neural network for detecting and classifying UAV audio from various environmental noises. This method converts UAV audio data into mel spectrograms and inputs them into the neural network model. The model automatically learns and identifies important time and frequency regions in the mel spectrogram through the attention mechanisms of time and frequency and assigns higher attention weights to them to improve recognition accuracy. Combined with soft thresholding，it suppresses the influence of environmental noise and outliers on the model and improves its classification and recognition performance under various environmental noise interferences. In this study，we collected an 8-category audio dataset for drones and enhanced it using background noise. We evaluated existing methods using this dataset. The results demonstrated that the proposed method is superior to existing methods in terms of accuracy，precision，recall，and F1-score.