Sound source localization (SSL) in noisy and reverberant environments remains a significant challenge, prompting extensive research into robust solutions. Recent advances in deep learning have demonstrated considerable potential for improving SSL performance. This paper presents a novel model that combines a convolutional neural network (CNN) with a gated recurrent unit (GRU) and a customized weighted feature fusion mechanism to enhance direction-of-arrival (DOA) estimation in acoustically complex conditions. The proposed CNN-GRU architecture extracts discriminative spatial features, including interaural time difference (ITD), interaural level difference (ILD), generalized cross-correlation phase transform (GCC-PHAT), and sound intensity from multichannel audio signals. To assess robustness and generalizability, two experiments are conducted, evaluating the model’s performance under varying additive noise levels and reverberation durations. Experimental results demonstrate that the proposed approach outperforms existing baseline methods, achieving a peak DOA estimation accuracy of 91.85% and a mean absolute error (MAE) of 2.46°, highlighting its effectiveness in challenging acoustic scenarios.