Urban planning refers to the efforts of designing land-use configurations. Effective urban planning can help to mitigate the operational and social vulnerability of a urban system, such as high tax, crimes, traffic congestion and accidents, pollution, depression, and anxiety. Due to the high complexity of urban systems, such tasks are mostly completed by professional planners. But, human planners take longer time. The recent advance of deep learning motivates us to ask: can machines learn at a human capability to automatically and quickly calculate land-use configuration, so human planners can finally adjust machine-generated plans for specific needs? To this end, we formulate the automated urban planning problem into a task of learning to configure land-uses, given the surrounding spatial contexts. To set up the task, we define a land-use configuration as a longitude-latitude-channel tensor, where each channel is a category of POIs and the value of an entry is the number of POIs. The objective is then to propose an adversarial learning framework that can automatically generate such tensor for an unplanned area. In particular, we first characterize the contexts of surrounding areas of an unplanned area by learning representations from spatial graphs using geographic and human mobility data. Second, we combine each unplanned area and its surrounding context representation as a tuple, and categorize all the tuples into positive (well-planned areas) and negative samples (poorly-planned areas). Third, we develop an adversarial land-use configuration approach, where the surrounding context representation is fed into a generator to generate a land-use configuration, and a discriminator learns to distinguish among positive and negative samples.