A multivariate spatiotemporal spread model of COVID-19 using ensemble of ConvLSTM networks

swarna kamal paul; Saikat Jana; Parama Bhaumik

The forecasting problem is framed as supervised learning problem. Given a sequence of observed matrices of spatial data as images 1 , 2 … the objective of the model is to predict the next image +1 . The training samples are divided into input sequences each of length W and output image. The model predicts the normalized log transformed new infections count in each pixel in a image. Thus, the output image consists of only 1 channel. The input training dataset (X train )can be represented as a tensor of size S train xWxLxLxC and the output dataset (Y train ) as S train x1xLxLx1. For training, the input sequences are selected from all frames having non-zero total infection count. Fig. 1b illustrates the sequence of images in a frame. The image t-6 to t-3 represents an input training sequence (X train ) of length W. The output image (Y train ) for this training sample . CC-BY 4.0 International license It is made available under a author/funder, who has granted medRxiv a license to display the preprint in perpetuity.