We introduce a U-Net model for 360◦ acoustic source local- ization framed as spherical semantic segmentation. Instead of regressing direction-of-arrival (DoA) angles, our model segments delay-and-sum (DAS) beamformed audio maps (azimuth × elevation) to highlight active sound regions. Us- ing a ground-based 24-microphone array, we recorded an open-field drone dataset synchronized with GPS flight logs for binary mask supervision. By operating on beamformed energy maps rather than raw audio, this hybrid approach uti- lizes pre-computed spatial representations to identify source distributions and suppress noise effectively. DoA estimates are derived via centroiding. Real-world experiments demon- strate improved angular precision over classical baselines across varying distances. Validation on the DCASE 2019 TAU Spatial dataset confirms our formulation generalizes beyond drones to multiclass Sound Event Localization and Detection (SELD).