Accurately identifying targets in the presence of noise remains a major challenge for Synthetic Aperture Radar (SAR) systems, particularly under low Signal-to-Noise Ratio (SNR) conditions. This study presents a Deep Learning (DL)-based framework for Single and Multiple Target Localization (SMTL) in SAR systems. A Linear Frequency Modulated Continuous Wave (LFMCW) signal model is formulated within a Compressive Sensing (CS) framework to enhance sparsity and resolution. The resulting Sample Covariance Matrix (SCM) is processed as a multi-channel input by a Convolutional Neural Network (CNN) to enable precise range estimation. The proposed model exhibits robust performance under severe noise conditions, achieving high localization accuracy even at SNR levels as low as −15 dB, thereby outperforming existing state-of-the-art methods in both accuracy and noise resilience.