Abstract: A mask-weighted deep learning sound source localization method is proposed for small-sized microphone arrays. Based on the power spectrum feature extraction, the co-located sound intensity feature in different directions is further integrated to solve the non-co-located problem in the sound intensity feature extraction, effectively improving the accuracy of time-frequency masking and the localization performance under reverberant and noisy environments. Furthermore, peak searching is not required, and low computational complexity is maintained. Simulation and experimental results show that the localization performance of proposed method outperforms the existing mask-weighted localization methods under reverberant and noisy environments, with an advantage becoming more pronounced under low signal-to-noise ratios and long reverberation times. The accuracy of sound source localization was found to exceed existing methods by 6.67%~10.00% in real-world experiments.