Multimodal Spatial-Temporal Prediction and Classification Using Deep Learning

“Spatial-temporal” (ST) data is time series data from multiple locations. Data is unpredictable, making predictions difficult. For a variety of urban tasks, such as estimating the speed of traffic and the demand for taxis, an accurate prediction that is based on this kind of data is required. Because of this assumption, the methods that are currently being used that are based on deep learning can only produce one possible outcome. As a consequence of this, they are unable to comprehend how the future will comprise a wide variety of components and how these components will interact with one another. Also, the current method operates under the presumption that information that is spatial and that which is temporal are essentially distinct, and as a result, each must be investigated separately. The chapter introduces a novel approach that utilises spatial-temporal convolutional neural networks in conjunction with Bi-LSTM and enhanced generative adversarial networks (E-GAN) to effectively capture non-linear correlations present in the data distribution. This is achieved through the use of inverse mapping from the forecast distribution. The spatio-temporal correlation network is a modelling technique that captures the distribution of pixels in both space and time. This approach enables the random sampling of latent variables to generate multiple future scenarios. This is accomplished by modelling the spatial distribution of pixels. This sampling can be carried out for a very wide variety of different possible outcomes (STCN). It is a stochastic adversarial network that learns to perform variational inference on data and generate data together with other people through implicit distribution modelling. Education is the means by which one can accomplish both of these goals. E-GAN also allows the combination of external factors, which further improves model learning, and it does this without any additional work. E-GAN outperforms the baseline models and significantly improves performance, as shown by extensive testing on two datasets derived from the real world.