<p>This paper presents an advanced automatic target recognition (ATR) system for ground surveillance radars (GSRs) that addresses the inherent complexity of raw radar data, including intricate Doppler and micro-Doppler signatures. The proposed solution is a novel multi-model deep learning framework that leverages a large, unique dataset acquired through simulation of different ground targets. This synthetic data was generated to model two distinct low-resolution radars operating in the Ku and X-bands. This framework incorporates a long short-term memory (LSTM) network for temporal sequences, a network that leverages audio analogies from the inverse fast Fourier transform (IFFT) of Doppler data, a convolutional neural network (CNN) for spectrogram images, and a network trained on statistical features. To enhance robustness, we employ diverse decision fusion strategies, including soft voting, hard voting, and a meta-classifier approach, to synergistically combine the outputs. A crucial element is the introduction of a novel dynamic temporal post-processing method, which we term temporal label smoothing (TLS). This method significantly improves robustness and accuracy by exploiting the temporal consistency of target behaviors. The system achieves on average 3 to 6% increase in accuracy over the best single-model classifier. The paper demonstrates substantial performance improvements across a wide array of diverse targets.</p>