Nowcasting is an important field in which weather predictions are required for a time scale less than one hour, but it still encounters some problems such as small changes and fluctuations in atmospheric conditions, coarse-grained data in space, and the dependency of the time series in the data model. Such a problem is a common difficulty when it comes to using traditional approaches to forecasting, especially when attempting to deal with significant fluctuations in specific weather parameters. Newly published findings in deep learning propose improved methodologies and mathematical algorithms for the interpretation of big weather data, for the extraction of detailed spatial/temporal features, and for real-time predictions. The purpose of this paper is to present a survey on the latest advancement of deep learning for short-term weather nowcasting in terms of model architecture, data source, and performance assessment. The focus is given to know-strengths-and-weaknesses of certain types of models such as CNNs, RNNs, and combined ones, as well as handling multiple data types among which satellite imagery data, radar data, and data from weather stations. This review aims at understanding the future work surrounding deep learning for weather nowcasting and envision the ways by which the capabilities of real-time weather forecasting can be improved in terms of accuracy and dependability.