<p>In recent years, there has been a surge focusing on advanced sound source localization methods based on deep learning. However, the black-box feature extraction mechanism impedes their optimization. This study proposes a two-step grid-free method for locating multiple sources based on deep ensemble learning and interpretable artificial intelligence (AI). In the first step, sound pressure signals received by a spherical microphone array are preprocessed into auto-power spectra, which are then fed into a classifier to count the sources. In the second step, the aforementioned signals are converted into generalized cross-correlations with phase transform (GCC-PHAT), which are then input to an ensemble regressor comprising four types of regressors for source localization. Additionally, two interpretable AI techniques, t-distributed stochastic neighbor embedding and activation maps, are employed to analyze the underlying principles of the deep learning-based method. The classifier achieves a testing accuracy of 94.59%. Compared to single-type regressors, the ensemble regressor reduces the mean localization error by up to 33.72%. The activation maps reveal that the extracted features are the middle section of the GCC-PHAT, while the influence of the two sides on localization is negligible. After trimming these sides, input data volume decreases by 72.05%, and the time per localization drops remarkably without significant loss in accuracy. These results demonstrate that the proposed method achieves three-dimensional multi-source localization with higher accuracy and lower computation cost.</p>