In modern wireless and sensing systems, effective communication and radar signal sensing are critical. Traditional methods, relying on engineered features or standalone neural networks, struggle with multi-modal data and low signal-to-noise ratios (SNR). They fail to capture spatial and temporal relationships and lack inter-modal connections. To overcome these issues, a novel CNN-LSTM framework has been proposed. It uses convolutional neural networks (CNN) for spatial feature extraction and long short- term memory networks (LSTM) to model temporal dependencies, refining features across modalities. By integrating IQ data, spectrograms, and cyclic spectrum representations, the model ensures robust signal sensing, even in challenging conditions. Experiments show the method achieves over 99.6% sensing accuracy in Gaussian and Rician channels and outperforms traditional approaches in low SNR environments, down to -5 dB. These results highlight the framework's potential to enhance communication and radar signal sensing. © 2025 IEEE.