Direction finding using spherical microphone array and spherical harmonics yields better accuracy compared to other methods. However, a major drawback in using the spherical microphone array is the full search in both polar and azimuth angle domains hence making it more computationally intensive. This project proposes a deep learning approach that classifies the spherical sector onto which the signal of interest arrives from. By identifying the spherical sector, the search for the beampattern peak is limited to the identified sector hence a significant reduction in the search space. The features for training a classifier is extracted from various layers of various deep learning networks. This will be followed by feature selection and finally classification using the selected features.