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

swarna kamal paul; Saikat Jana; Parama Bhaumik

Experiments have been carried out to predict the future new infection cases in Italy for a period of 5 days and 10 days and in USA for a period of 5 days and 8 days. Data has been collected from Harvard dataverse [15, 16] and [19] . For USA the data collection period is '2020-03-09' to '2020-04-08' and for Italy it is '2020-02-05' to '2020-04-10'. Test data period for Italy data is '2020-04-01' to '2020-04-10' and for USA it is '2020-04-01' to '2020-04-08'. Fig. 2a shows the region of USA which has been divided into 18x30 grids. The length W of each training input sequence is taken as 10 days. As shown in Fig. 2b , the region of Italy is divided in 7x6 grids. For both the countries the frames containing at least a single Covid-19 infection case will be considered for training and testing the model. The Covid-19 cases are marked in red bubbles in the map. Each frame in turn is divided in 16x16 pixels with an overlap Margin of 4 pixel. Training sequences containing at least one positive case are only selected. The model consists of ensemble of 5 Convolutional LSTM networks. For Italy each network contains 4 hidden layers with sigmoid activation. The output layer is a Convolutional 3D layer with exponential linear unit as activation. The models are trained for 30 epochs with mean squared error as loss function. Each Conv2D layer has kernel of size 3x3. The Conv3D layer has kernel size 3x3x3. The input and hidden layers have 32 filters. The input layer is configured to take images of size 16x16x2. Each frame in the input sample have 2 channels for 2 features. The first channel contains the normalized log transformed count of new cases per day. The second channel contains the normalized log transformed population in each pixel with no temporal variation. The region in USA is approximately 13 times than Italy and the distribution of Covid-19 cases in USA is geospatially highly skewed. Thus, grids are divided in four equal sections by a latitude and . 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.