We introduce a method for detecting, localizing and identifying radio transmissions within wide-band time-frequency power spectrograms using feature learning using convolutional neural networks on their 2D image representation. By doing so we build a foundation for higher level contextual radio spectrum event understanding, labeling, and reasoning in complex shared spectrum and many-user environments by developing tools which can rapidly understand and label sequences of events based on experience and labeled data rather than signal-specific detection algorithms such as matched filters.