Convolutional neural networks with hierarchical context transfer for high-resolution spatiotemporal predictions

In this work, we consider a problem of predicting the next state of a given area using retrospective information. The proposed concept of hierarchical context transfer (HCT) operates on several spatial levels of the input data to overcome major issues of next state prediction problems - density variability, a significant difference between consecutive states and computational complexity. The custom loss function allows assimilating contexts of spatial levels into each other to further improve prediction quality. The proposed deep learning model (HCT-CNN) allows generating precise high-resolution predictions of the target area. We evaluate our model on the use case of predicting the next state of the urban area using a large dataset for six cities - New York, Moscow, London, Tokyo, Saint Petersburg, and Vienna. Experimental results demonstrate that HCT-CNN generates low- and high-resolution predictions of better quality than existing methods.